[SOT] add direct kernel fast runtime

[SigureMo]

PR Category

Performance Optimization

PR Types

New features, Performance

Description

本 PR 为 SOT 动转静执行链路增加 fast kernel runtime，用于压缩运行时调度开销。核心目标是绕过原 C++ program executor 的逐 op 调度路径，在 PIR lower 到 pd_kernel dialect 后直接 codegen Python 可执行函数，并通过自动生成的 core.eager.kernel_ops pybind API 调用底层 kernel。

主要改动：

• 新增 pir_fast_runtime.py，从 lower 后的 pd_kernel/CINN IR 生成 Python runtime code。
• SOT PartialProgramLayer.sot_call 增加可开关的 fast kernel runtime 路径。
• 新增 direct kernel pybind API 生成与 metadata 查询能力，覆盖普通 PHI kernel 调用。
• 新增 CINN JIT kernel launch 的 Python 入口，用于 GPU 下 lower 后的 CINN kernel 调用。
• feed/parameter 直接绑定到 Python locals，不额外创建 Scope。
• 对 unsupported op 直接报错，不 fallback 到 run_program，避免隐藏问题。

验证

• prek --files python/paddle/jit/dy2static/pir_fast_runtime.py test/sot/test_fast_kernel_runtime.py
• python -m unittest test_fast_kernel_runtime
• ctest -R sot --output-on-failure
• 远端 GPU build 完成，并通过 core.eager.kernel_ops.add smoke test 与 test_fast_kernel_runtime。
• GitHub CI 中 PR-CI-SOT / Build and Test 已通过。

性能结果

以下 benchmark 均排除首次 codegen/CINN compile 时间。

GPU：

• single_add：sync 22.866us -> 19.605us，enqueue 18.336us -> 14.357us。
• reshape_add：sync 60.321us -> 53.518us，enqueue 55.852us -> 48.256us。
• resnet18_b1：sync 0.930ms -> 0.800ms，enqueue 0.824ms -> 0.739ms。
• resnet18_b10：sync 0.999ms -> 0.855ms，enqueue 0.912ms -> 0.813ms。

CPU：

• single_add：3.589us -> 1.501us。
• reshape_add：18.617us -> 5.620us。
• resnet18_b1：8.627ms -> 8.096ms。
• resnet18_b10：78.468ms -> 76.781ms。

是否引起精度变化

否。该 PR 仅调整 SOT 编译图运行时调度路径，kernel 输入输出保持与 lower 后 IR 一致；新增单测覆盖了 direct kernel、CINN kernel、BatchNorm eval 和禁止 run_program fallback 的行为。

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[Copilot]

Pull request overview

This PR introduces a "direct kernel fast runtime" path for SOT (Symbolic OpTracing) in Paddle. When the new SOT_ENABLE_FAST_KERNEL_CODEGEN environment variable is enabled, sot_call lowers the PIR program to the pd_kernel dialect and code-generates a Python function that directly invokes per-op pybind kernels (newly exposed under core.eager.kernel_ops), bypassing the run_program executor. The path intentionally does not fall back to run_program and raises a RuntimeError if autograd is active or required metadata is missing.

Changes:

• Adds a codegen mode (--direct_kernel) to python_c_gen.py plus new CMake wiring that produces kernel_op_function.{h,cc} exposing direct kernel APIs and a get_kernel_ops_args_info registry under core.eager.kernel_ops.
• Adds manual pybind helpers in kernel_op_function_manual.{h,cc} (incl. get_phi_kernel_op_info and a run_cinn_jit_kernel launcher), exposes apply_pd_op_to_kernel_pass in pir.cc, and binds the new submodule from eager.cc.
• Adds pir_fast_runtime.py implementing the lowering + Python source codegen, plumbs it into PirPartialProgram.sot_call via a cached FastKernelRuntime, and adds unit tests in test/sot/test_fast_kernel_runtime.py.

Reviewed changes

Copilot reviewed 13 out of 14 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
python/paddle/jit/sot/utils/envs.py	Adds SOT_ENABLE_FAST_KERNEL_CODEGEN boolean env flag.
python/paddle/jit/dy2static/pir_partial_program.py	Routes sot_call to the fast kernel path when enabled; adds runtime cache and quick_index_map handling for Value raw inputs.
python/paddle/jit/dy2static/pir_fast_runtime.py	New module that lowers PIR to pd_kernel and codegens a Python kernel-dispatch function with constant folding and CINN jit_kernel support.
paddle/fluid/pybind/pir.cc	Exposes apply_pd_op_to_kernel_pass to Python.
paddle/fluid/pybind/eager.{h,cc}	Declares and creates the core.eager.kernel_ops submodule and binds both generated and manual direct-kernel functions.
paddle/fluid/pybind/kernel_op_function_manual.{h,cc}	Implements get_phi_kernel_op_info and a CINN jit_kernel launcher used by the fast runtime.
paddle/fluid/pybind/CMakeLists.txt, .gitignore, paddle/fluid/pir/dialect/CMakeLists.txt	Build wiring for generated kernel_op_function.* files.
paddle/fluid/eager/auto_code_generator/generator/python_c_gen.py	Extends the Python-C generator with a --direct_kernel mode, custom namespaces, args-info registry, and configurable bind/method/error names.
test/sot/test_fast_kernel_runtime.py	Unit tests covering pybind surface, fast kernel codegen for add/reshape/BN, and the no-fallback contract.
.gitignore	Ignores generated kernel_op_function.* outputs.

---

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

[codecov-commenter]

Codecov Report

❌ Patch coverage is 64.37346% with 145 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@3fc714d). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
python/paddle/jit/dy2static/pir_fast_runtime.py	62.71%	132 Missing ⚠️
python/paddle/jit/dy2static/pir_partial_program.py	75.00%	13 Missing ⚠️

❌ Your patch status has failed because the patch coverage (64.37%) is below the target coverage (90.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop   #79043   +/-   ##
==========================================
  Coverage           ?   64.37%           
==========================================
  Files              ?        3           
  Lines              ?      407           
  Branches           ?        0           
==========================================
  Hits               ?      262           
  Misses             ?      145           
  Partials           ?        0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.