PID FOPDT full-chain E2E harness

End-to-end validation of the FOPDT PID-tuning pipeline, chaining the real operators fit → tuning_rule → fopdt_rollout → performance_metrics and comparing against a CPU reference.

Two tools are provided:

Tool	Purpose	Compares against
e2e_orchestrator.py	Accuracy: drives the 4 operators stage-by-stage (e2e_runner) and checks each stage against its Python reference.	per-stage CPU reference (common/*_reference.py)
e2e_perf	Performance: single-process, device-resident chain tuning_rule → fopdt_rollout → performance_metrics, timed vs a CPU 64-thread chain; also re-checks final best-PID / score / metrics alignment.	CPU multi-thread chain (in-process)

The rollout stage dominates the chain cost (tuning/metrics are ~0.05 ms each), so the chain speedup tracks the rollout speedup.

Build

The operators must be built first (each <op>/build/lib<op>_host.so and <op>/build/lib/lib<op>_kernel_lib.so present). Then, from this directory:

bash build_e2e.sh        # produces ./e2e_perf and ./e2e_runner

Override the toolkit location with ASCEND_HOME / ASCEND_TOOLKIT_ENV if it is not at the default /usr/local/Ascend/ascend-toolkit/latest.

Run — performance (e2e_perf)

# args: <device> [batch=128] [candidates=1024] [sim_steps=1024] \
#       [candidate_tile=0:auto] [iters=5] [warmup=2] [threads=64]
./e2e_perf 0 128 16384 1024 0 5 2 64

candidate_tile=0 lets the rollout operator auto-select the optimal tile (min(candidates, kLane=768)); pass an explicit value only to sweep the knob. Example representative-scale result (Ascend910B3, B=128, sim_steps=1024, auto tile): C=1024 ≈ 4.0x, C=4096 ≈ 6.2x, C=16384 ≈ 4.5x vs CPU 64T.

Run — accuracy (e2e_orchestrator.py)

export E2E_RUNNER=$PWD/e2e_runner   # required: path to the built runner
export E2E_WORK=/tmp/e2e_work       # optional: scratch dir for .bin I/O
# export PID_COMMON=/path/to/PIDModelFit/common   # optional override; defaults to ../common
python3 e2e_orchestrator.py

It writes a per-stage comparison report to $E2E_WORK/e2e_report.json and prints the max error of each stage (NPU vs reference). All four stages align to within float32 tolerance.