Update threading configuration for NVL and PTL

[sunxiaoxia2022]

Details:

• Update threading configuration for NVL and PTL

1. Improve hybrid CPU latency heuristics to enable more ALL/AUTO configurations on high LP E-core-share platforms.
2. Add new model-profile-based AUTO rules using lp_ecore_share, memory tolerance, convolution pressure, and GEMM ratio.
3. Refine latency stream sizing so model_prefer_threads is honored when building single-stream ALL core configurations.
4. Extend CPUStreamsExecutor internal stream metadata from a single core type to multiple core types.
5. Update task arena creation logic to bind streams by the actual set of eligible core types instead of forcing a single-type binding.
6. Add unit coverage for mixed-core stream metadata and stream type selection behavior.

Dependency PR: PR#36286

Tickets:

• CVS-183947
• CVS-180761

Test data:

Machine:
1.NVL windows from validation team (8Pcores+16Ecores+4LPEcores),
2.PTL X7 358H 32GB windows (4Pcores+8Ecores+4LPEcores)
Models: 209 static models from WW05_static_2026.0.0-20947

The geomean of latency with PR vs master:
NVL: 0.97583
PTL: 0.92095

The performance data for all models:

The performance data for the models with threads and partitioner changed in current PR:

[Copilot]

Pull request overview

Updates CPU threading heuristics and stream binding metadata to better support hybrid-core platforms (NVL/PTL), including allowing streams to be associated with multiple eligible core types and refining AUTO/LATENCY stream sizing behavior.

Changes:

• Extend stream metadata from a single core_type to a list of core_types, and update task arena creation to bind based on the eligible core-type set.
• Add new model/memory-pressure-based heuristics that select ALL-core AUTO configurations more often on high LP E-core-share systems.
• Expand unit coverage for streams-info table cases, model-prefer-threads rules, and mixed-core stream binding behavior.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/plugins/intel_cpu/tests/unit/streams_info/streams_info_table_test.cpp	Adds a new LATENCY test case covering an ALL+MAIN+EFFICIENT+LP_EFFICIENT stream table scenario.
src/plugins/intel_cpu/tests/unit/streams_info/model_prefer_threads.cpp	Adds multiple integration tests validating new hybrid AUTO selection and stream table expectations.
src/plugins/intel_cpu/src/cpu_streams_calculation.cpp	Introduces new heuristic profiles for selecting AUTO/ALL-core latency behavior, plus stream sizing refinement.
src/inference/tests/unit/cpu_stream_info_test.cpp	Updates stream-info unit tests for multi-core-type metadata and adds new coverage cases.
src/inference/src/dev/threading/cpu_streams_executor.cpp	Updates task arena creation to handle multiple core types for binding and threads-per-core behavior.
src/inference/src/dev/threading/cpu_streams_executor_internal.cpp	Updates stream-info extraction to return a list of core types and selects binding type based on subset eligibility.
src/inference/dev_api/openvino/runtime/threading/cpu_streams_executor_internal.hpp	Updates internal API signature/docs to return core_types instead of a single core_type.

[Copilot]

Pull request overview

Copilot reviewed 7 out of 7 changed files in this pull request and generated 6 comments.

[Copilot]

Pull request overview

Copilot reviewed 7 out of 7 changed files in this pull request and generated 3 comments.