PidResidualDiagnostics Benchmark Report【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测试环境NPU 环境node202SOCAscend910B3NPU device3CANN/usr/local/Ascend/ascend-toolkitCPU 对比线程数64数据类型float32测试口径cpu_parallelCPU 64 线程计算 8 个残差指标和max_lag个自相关。npu_kernel输入已在 Device仅统计算子执行时间。npu_e2e包含actual/predictedH2D、kernel、metrics/autocorrD2H。当前输出指标mean_residual,std_residual,mae,rmse,max_abs_residual,fit_percent,durbin_watson,ljung_box_q当前版本固定输出 8 个指标使用fit_percent作为拟合优度指标。数值和性能结论均基于该 8 指标版本。正确性SmokePidResidualDiagnostics smoke mean0 mae0.5 rmse0.707107 max_abs1 dw1.5 autocorr[0, -0.5] PASSEDBenchmark 中 CPU/NPU 输出对比max_autocorr_abs_err 0max_metric_abs_err 0.0054最大 metric 误差来自fit_percent相对误差约5.8e-5性能结果BNmax_lagCPU 64T msNPU kernel msNPU e2e mskernel/CPU64Te2e/CPU64T1281024160.4689490.0920590.4426515.09x1.06x2561024321.399990.05525260.35978225.34x3.89x5122048322.333110.0757710.95757330.79x2.44x10242048648.628030.06830551.85133126.31x4.66x结论PidResidualDiagnostics是本次正式提交的模型诊断算子之一业务价值明确用于 PID 模型辨识后的残差白噪声检查、拟合质量评估和模型可信度判断。计算结构适合 NPU批量归约、多 lag 相关性扫描、输出小。e2e 也能超过 CPU 64 线程说明即使输入来自 Host该算子在中大规模下仍有实际加速空间。复现命令source /usr/local/Ascend/ascend-toolkit/set_env.sh cd prediction/ProcessControl/PIDModelFit/pid_residual_diagnostics cmake -S . -B build -DCMAKE_BUILD_TYPERelease -DSOC_VERSIONAscend910B3 cmake --build build -j 2 export LD_LIBRARY_PATH$PWD/build:$PWD/build/lib:/usr/local/Ascend/ascend-toolkit/latest/lib64:/usr/local/Ascend/ascend-toolkit/8.1/lib64:${LD_LIBRARY_PATH:-} ./build/test_aclnn_pid_residual_diagnostics 3 ./build/benchmark_pid_residual_diagnostics 3 128 1024 16 5 1 64 ./build/benchmark_pid_residual_diagnostics 3 256 1024 32 5 1 64 ./build/benchmark_pid_residual_diagnostics 3 512 2048 32 3 1 64 ./build/benchmark_pid_residual_diagnostics 3 1024 2048 64 2 1 64【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创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考