Weird Events

关注

原创最新推荐文章于 2026-10-12 16:29:44 发布 608 阅读 AI 写同款· GEO 优化 ›

【信息科学与工程学】计算机科学与自动化——第三百零五篇数据中心 Scale-Up、Scale-Out、Scale-Across 20 GPU 训练中的各种算子融合技巧、梯度压缩与稀疏通信的高级变体、混合专家模型的负载均衡与通信优化、长序列模型的注意力机制优化、模型量化与蒸馏的联合优化、分布式训练中的容错与恢复、以及一些新兴硬件特性的建模（如 HBM3、GDDR7、CXL 3.0 内存语义、PCIe Gen6 PAM4 信令等）。每条仍然包含四个方程式子节。阅读详情

Q1: (surrendered) istream::gcount() can not work?
A1: Now I apply strlen() function to the read-buffer instead of invocation to gcount() as a temporary alternative.
I should overview and debug the STL and Poco source code to find out what caused this event happened,
but time is not too much for me now, I promise here that I'll be back later when I have time. o(^_^)o

Q2: (cleared) Due to low-level implementation, Poco::FastMutex is recursive on win32 platform, and non-recursive on posix platform.
A2: CRITICAL_SECTION is default to be recursive, therefore FastMutex who just wrapping the EnterCriticalSection() call and do nothing else is not that adequate when whose code comment told me that FastMutex run as a non-recursive lock.

Q3: (cleared) Deadlock incurred in the scope with a scope-lock wrapper whose destructor when leave the scope will release the private mutex that acquired in the constructor at initial point of the scope.
A3: I admit my fault. -_-! I make potential lock-acquire call in the scope even who really does have a scope-lock wrapper.

收录于