Support: add l2-orchestrator-standalone bench for the L2 submit path

l2-orchestrator-standalone/.gitignore

@@ -0,0 +1,4 @@
+# Every build tree this package makes. scripts/profile_l2.sh alone creates
+# three (build, build-prof, build-tm), and the root .gitignore's `build/` rule
+# matches only the first.
+build*/

l2-orchestrator-standalone/CMakeLists.txt

@@ -0,0 +1,156 @@
+# Standalone L2 orchestrator package: bench + profile for the four-line
+# orchestration sequence in runtime_maker.cpp:538-541. Host-only, no CANN,
+# no Ascend SDK, no NPU.
+#
+# The TU list mirrors the `host` target of
+# src/a2a3/runtime/host_build_graph/build_config.py, minus `host/` itself
+# (that is where runtime_maker and its CANN dependencies live):
+#
+#   BUILD_CONFIG["host"]["source_dirs"] =
+#       ["host", "runtime/orchestrator_core", "runtime/shared", "orchestration"]
+#
+# scripts/check_extraction.sh asserts this list still matches the source repo.
+
+cmake_minimum_required(VERSION 3.15)
+# C is required, and not incidentally: qwen3_dynamic_tensormap.h is C and cannot
+# be compiled as C++ (C99 compound-literal lvalues; see esl_shim/esl_c_abi.h).
+project(l2_orchestrator CXX C)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_CXX_EXTENSIONS OFF)
+set(CMAKE_C_STANDARD 11)
+set(CMAKE_C_STANDARD_REQUIRED ON)
+
+if(WIN32)
+    message(FATAL_ERROR "This package requires POSIX. Build on Linux.")
+endif()
+
+# Neither simpler nor this package sets a build type by default, which would
+# leave the bench measuring an -O0 engine. Pin it: an unoptimised orchestrator
+# is not the thing anyone wants a number for.
+if(NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "" FORCE)
+endif()
+
+set(ROOT ${CMAKE_CURRENT_SOURCE_DIR})
+# Engine TUs are compiled straight out of the enclosing checkout, not copied:
+# the bench measures the orchestrator of THIS revision, and an edit to
+# ../src is picked up by the next build with nothing to keep in sync.
+set(S ${ROOT}/../src)
+set(R ${S}/a2a3/runtime/host_build_graph)
+
+# SIMPLER_ORCH_PROFILING gates the engine's own per-STEP cycle counters inside
+# submit_task_common (Level 3 of the profile report). It requires SIMPLER_DFX=1,
+# and both add work to the measured path — so the default is off and the
+# throughput numbers come from the clean build.
+option(L2_ORCH_PROFILING "Build the engine's own per-STEP cycle counters (Level 3)" OFF)
+# SIMPLER_TENSORMAP_PROFILING adds the engine's own TensorMap lookup counters
+# (bucket chain length, overlap checks/hits, insert count) — Level 4 of the
+# report. profiling_config.h #errors unless SIMPLER_ORCH_PROFILING is also on,
+# so this implies it.
+option(L2_TENSORMAP_PROFILING "Build the engine's TensorMap lookup counters (Level 4)" OFF)
+
+add_library(l2_engine STATIC
+    ${R}/runtime/orchestrator_core/pto_orchestrator.cpp
+    ${R}/runtime/orchestrator_core/pto_ring_buffer.cpp
+    ${R}/runtime/orchestrator_core/pto_runtime2.cpp
+    ${R}/runtime/shared/pto_runtime2_init.cpp
+    ${R}/runtime/shared/pto_shared_memory.cpp
+    ${R}/runtime/shared/pto_tensormap.cpp
+    ${R}/runtime/shared/runtime.cpp
+    ${R}/orchestration/common.cpp
+    ${ROOT}/src/host_shim/host_shim.cpp
+)
+target_include_directories(l2_engine PUBLIC
+    ${R}/runtime
+    ${R}/common
+    ${R}/orchestration
+    ${S}/a2a3/runtime
+    ${S}
+    ${S}/common
+    ${S}/common/log/include
+    ${S}/common/task_interface
+    ${S}/common/platform/include
+    ${S}/a2a3/platform/include
+)
+if(L2_TENSORMAP_PROFILING)
+    set(L2_ORCH_PROFILING ON)
+endif()
+if(L2_ORCH_PROFILING)
+    target_compile_definitions(l2_engine PUBLIC SIMPLER_DFX=1 SIMPLER_ORCH_PROFILING=1)
+    message(STATUS "SIMPLER_ORCH_PROFILING=1 — Level 3 built; this build is NOT the one to quote throughput from")
+endif()
+if(L2_TENSORMAP_PROFILING)
+    target_compile_definitions(l2_engine PUBLIC SIMPLER_TENSORMAP_PROFILING=1)
+    message(STATUS "SIMPLER_TENSORMAP_PROFILING=1 — Level 4 (TensorMap chain stats) built")
+endif()
+
+# Where qwen3_dynamic_tensormap.h lives. Defaults to this package's own
+# directory; override to point elsewhere.
+set(L2_QWEN3_DYN_CASE_DIR "${ROOT}" CACHE PATH
+    "Directory containing qwen3_dynamic_tensormap.h")
+if(NOT EXISTS "${L2_QWEN3_DYN_CASE_DIR}/qwen3_dynamic_tensormap.h")
+    message(FATAL_ERROR
+        "qwen3_dynamic_tensormap.h not found in ${L2_QWEN3_DYN_CASE_DIR}.\n"
+        "Pass -DL2_QWEN3_DYN_CASE_DIR=<dir> to point at it.")
+endif()
+
+# The payload is separate from the engine so a bench-side change cannot reach
+# the engine TUs.
+add_library(l2_payloads STATIC
+    ${ROOT}/bench/payload_qwen3_dyn.cpp
+    ${ROOT}/bench/payload_qwen3_dyn_case.c      # C, on purpose — see below
+    ${ROOT}/bench/esl_shim/esl_shim_impl.cpp
+)
+target_include_directories(l2_payloads PUBLIC ${ROOT}/bench)
+target_link_libraries(l2_payloads PUBLIC l2_engine)
+
+# esl_shim/ is deliberately NOT on any target-wide include path: it provides
+# headers named `mem_pool.h` and `tensormap.h`, and only the two TUs below may
+# see them. Everything else in the package must keep resolving those names the
+# way it does today (i.e. not at all).
+set_source_files_properties(${ROOT}/bench/esl_shim/esl_shim_impl.cpp PROPERTIES
+    INCLUDE_DIRECTORIES "${ROOT}/bench/esl_shim"
+)
+
+# The case's SPMD tier is a compile-time knob of the case itself:
+#
+#     #ifndef QWEN3_SPMD_TIER
+#     #define QWEN3_SPMD_TIER 4
+#     #endif
+#
+# and it changes the DAG, not just its scheduling. `qwen3_blocks_per_task()`
+# returns min(total_chunks, {1,2,4,8,1<<30}[tier]), so the tier decides how many
+# SPMD chunks are folded into ONE task. This package pins tier 0 — one chunk per
+# task, 3096 tasks / 3096 subtasks, the most task-dense variant and therefore the
+# one that puts the most pressure on the orchestrator per unit of device work.
+# The #ifndef guard in the case exists precisely so the build can set it, which
+# is why pinning it here does not modify the case file.
+#
+# The entry symbol is renamed at the same time so the case's C-linkage
+# `aicpu_orchestration_entry` cannot collide with the engine-side name.
+set_source_files_properties(${ROOT}/bench/payload_qwen3_dyn_case.c PROPERTIES
+    INCLUDE_DIRECTORIES "${L2_QWEN3_DYN_CASE_DIR};${ROOT}/bench/esl_shim"
+    COMPILE_DEFINITIONS "aicpu_orchestration_entry=qwen3_dyn_orchestration_entry;QWEN3_SPMD_TIER=0"
+)
+
+add_executable(l2_bench ${ROOT}/bench/l2_bench.cpp ${ROOT}/bench/l2_profile.cpp)
+target_link_libraries(l2_bench PRIVATE l2_payloads)
+
+add_executable(l2_orch_main ${ROOT}/apps/l2_orch_main.cpp)
+target_include_directories(l2_orch_main PRIVATE ${ROOT}/bench)
+target_link_libraries(l2_orch_main PRIVATE l2_payloads)
+
+enable_testing()
+add_test(NAME smoke COMMAND l2_orch_main)
+# Both entry points are covered because the throughput path and the profile path
+# drive the entry differently (the latter through the replacement ops table).
+add_test(NAME bench COMMAND l2_bench --mode=throughput --repeat=1)
+add_test(NAME profile COMMAND l2_bench --mode=profile)
+# The prefault diagnostics touch the SM and arena before the engine initialises
+# them, so a regression there would corrupt the run rather than just skew a
+# number. Cover both, and assert the graph still comes out the same size.
+add_test(NAME prefault COMMAND l2_bench --mode=throughput --repeat=1 --prefault-all)
+set_tests_properties(prefault PROPERTIES
+    PASS_REGULAR_EXPRESSION "kernel submits +3096 +framework allocs +779")