CANN PID整定全链路端到端验证

PID FOPDT full-chain E2E harness

【免费下载链接】mat-chem-sim-pred面向工业领域，聚焦计算仿真、预测两大核心场景，构建面向流程工业"机理+数据"双轮驱动的领域计算层，推动AI for Science在材料化学领域的深度应用。项目地址: https://gitcode.com/cann/mat-chem-sim-pred

End-to-end validation of the FOPDT PID-tuning pipeline, chaining the real operatorsfit → tuning_rule → fopdt_rollout → performance_metricsand comparing against a CPU reference.

Two tools are provided:

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.soand<op>/build/lib/lib<op>_kernel_lib.sopresent). Then, from this directory:

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

Override the toolkit location withASCEND_HOME/ASCEND_TOOLKIT_ENVif 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=0lets 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.jsonand prints the max error of each stage (NPU vs reference). All four stages align to within float32 tolerance.

【免费下载链接】mat-chem-sim-pred面向工业领域，聚焦计算仿真、预测两大核心场景，构建面向流程工业"机理+数据"双轮驱动的领域计算层，推动AI for Science在材料化学领域的深度应用。项目地址: https://gitcode.com/cann/mat-chem-sim-pred