Skip to content

4.1 Unified Memory ​

待翻译

本节已建立目录框架,正文尚未翻译。

4.1.1 Unified Memory on Devices with Full CUDA Unified Memory Support ​

4.1.1.1 Unified Memory: In-Depth Examples ​

4.1.1.1.1 File-backed Unified Memory ​

4.1.1.1.2 Inter-Process Communication (IPC) with Unified Memory ​

4.1.1.2 Performance Tuning ​

4.1.1.2.1 Memory Paging and Page Sizes ​

4.1.1.2.1.1 Choosing the Right Page Size ​

4.1.1.2.1.2 CPU and GPU Page Tables: Hardware Coherency vs. Software Coherency ​

4.1.1.2.1.3 Mixing Hardware and Software Coherency ​

4.1.1.2.2 Direct Unified Memory Access from the Host ​

4.1.1.2.3 Host Native Atomics ​

4.1.1.2.4 Atomic Accesses and Synchronization Primitives ​

4.1.1.2.5 Memcpy()/Memset() Behavior With Unified Memory ​

4.1.1.2.6 Overview of Memory Allocators for Unified Memory ​

4.1.1.2.7 Locality Domains and Unified Memory ​

4.1.1.2.8 Access Counter Migration ​

4.1.1.2.9 Avoid Frequent Writes to GPU-Resident Memory from the CPU ​

4.1.1.2.10 Exploiting Asynchronous Access to System Memory ​

4.1.2 Unified Memory on Devices with only CUDA Managed Memory Support ​

4.1.3 Unified Memory on Windows, WSL, and Tegra ​

4.1.3.1 Multi-GPU ​

4.1.3.2 Coherency and Concurrency ​

4.1.3.3 Stream Associated Unified Memory ​

4.1.3.3.1 Stream Callbacks ​

4.1.3.3.2 Managed memory associated to streams allows for finer-grained control ​

4.1.3.3.3 A more elaborate example on multithreaded host programs ​

4.1.3.3.4 Data Movement of Stream Associated Unified Memory ​

4.1.4 Performance Hints ​

4.1.4.1 Data Prefetching ​

4.1.4.2 Data Usage Hints ​

4.1.4.3 Memory Discarding ​

4.1.4.4 Querying Data Usage Attributes on Managed Memory ​

4.1.4.5 GPU Memory Oversubscription ​