https://blog.youkuaiyun.com/Beforthingzh-cnhttps://blog.youkuaiyun.com/https://blog.youkuaiyun.com/Beforthing/article/details/170366158https://blog.youkuaiyun.com/Beforthing/article/details/170366158BeforthingMon, 09 Nov 2026 20:38:26 +0800https://blog.youkuaiyun.com/Beforthing/article/details/170366114https://blog.youkuaiyun.com/Beforthing/article/details/170366114BeforthingMon, 09 Nov 2026 20:37:13 +0800https://blog.youkuaiyun.com/Beforthing/article/details/170366058https://blog.youkuaiyun.com/Beforthing/article/details/170366058BeforthingMon, 09 Nov 2026 20:35:50 +0800https://blog.youkuaiyun.com/Beforthing/article/details/169952081https://blog.youkuaiyun.com/Beforthing/article/details/169952081BeforthingTue, 03 Nov 2026 22:21:43 +0800https://blog.youkuaiyun.com/Beforthing/article/details/169952006https://blog.youkuaiyun.com/Beforthing/article/details/169952006BeforthingTue, 03 Nov 2026 22:16:06 +0800 BCELoss，避免数值溢出训练循环顺序：清零梯度 → 前向计算loss → backward → step。]]>https://blog.youkuaiyun.com/Beforthing/article/details/169155310https://blog.youkuaiyun.com/Beforthing/article/details/169155310BeforthingTue, 03 Nov 2026 22:07:51 +0800https://blog.youkuaiyun.com/Beforthing/article/details/168976583https://blog.youkuaiyun.com/Beforthing/article/details/168976583BeforthingMon, 26 Oct 2026 12:40:12 +0800https://blog.youkuaiyun.com/Beforthing/article/details/168976540https://blog.youkuaiyun.com/Beforthing/article/details/168976540BeforthingSat, 24 Oct 2026 19:38:10 +0800https://blog.youkuaiyun.com/Beforthing/article/details/168976531https://blog.youkuaiyun.com/Beforthing/article/details/168976531BeforthingSat, 24 Oct 2026 19:37:11 +0800https://blog.youkuaiyun.com/Beforthing/article/details/168976514https://blog.youkuaiyun.com/Beforthing/article/details/168976514BeforthingSat, 24 Oct 2026 19:35:54 +0800https://blog.youkuaiyun.com/Beforthing/article/details/167392217https://blog.youkuaiyun.com/Beforthing/article/details/167392217BeforthingFri, 09 Oct 2026 17:11:19 +0800https://blog.youkuaiyun.com/Beforthing/article/details/167391998https://blog.youkuaiyun.com/Beforthing/article/details/167391998BeforthingFri, 09 Oct 2026 17:05:30 +0800https://blog.youkuaiyun.com/Beforthing/article/details/167391769https://blog.youkuaiyun.com/Beforthing/article/details/167391769BeforthingFri, 09 Oct 2026 17:00:34 +0800https://blog.youkuaiyun.com/Beforthing/article/details/167391447https://blog.youkuaiyun.com/Beforthing/article/details/167391447BeforthingFri, 09 Oct 2026 16:52:35 +0800https://blog.youkuaiyun.com/Beforthing/article/details/167390735https://blog.youkuaiyun.com/Beforthing/article/details/167390735BeforthingFri, 09 Oct 2026 16:40:51 +0800 共享内存 > L1/L2 > 全局显存，访存是AI算力最大瓶颈；GPU AI编程本质：CPU负责调度，GPU Kernel定义单线程任务，启动海量线程并行；AX+Y（AXPY）是理解CUDA并行的入门样板，所有矩阵运算、神经网络算子都建立在这个并行思想之上。]]>https://blog.youkuaiyun.com/Beforthing/article/details/164066951https://blog.youkuaiyun.com/Beforthing/article/details/164066951BeforthingWed, 26 Aug 2026 00:03:29 +0800 PM‑QoS延迟约束 > Wakelock唤醒锁> CPUFreq / CPUIdle / Devfreq调度决策> Runtime‑PM单外设休眠> System‑Suspend整机休眠。]]>https://blog.youkuaiyun.com/Beforthing/article/details/164066905https://blog.youkuaiyun.com/Beforthing/article/details/164066905BeforthingWed, 26 Aug 2026 00:03:05 +0800https://blog.youkuaiyun.com/Beforthing/article/details/164066799https://blog.youkuaiyun.com/Beforthing/article/details/164066799BeforthingTue, 25 Aug 2026 23:47:22 +0800https://blog.youkuaiyun.com/Beforthing/article/details/164066776https://blog.youkuaiyun.com/Beforthing/article/details/164066776BeforthingTue, 25 Aug 2026 23:42:00 +0800https://blog.youkuaiyun.com/Beforthing/article/details/164066736https://blog.youkuaiyun.com/Beforthing/article/details/164066736BeforthingTue, 25 Aug 2026 23:37:08 +0800