Skip to content

3.5 A Tour of CUDA Features ​

待翻译

本节已建立目录框架,正文尚未翻译。

3.5.1 Improving Kernel Performance ​

3.5.1.1 Asynchronous Barriers ​

3.5.1.2 Asynchronous Data Copies and the Tensor Memory Accelerator (TMA) ​

3.5.1.3 Pipelines ​

3.5.1.4 Work Stealing with Cluster Launch Control ​

3.5.2 Improving Latencies ​

3.5.2.1 Green Contexts ​

3.5.2.2 Locality Domains ​

3.5.2.3 Stream-Ordered Memory Allocation ​

3.5.2.4 CUDA Graphs ​

3.5.2.5 Programmatic Dependent Launch ​

3.5.2.6 Lazy Loading ​

3.5.3 Functionality Features ​

3.5.3.1 Extended GPU Memory ​

3.5.3.2 Dynamic Parallelism ​

3.5.4 CUDA Interoperability ​

3.5.4.1 CUDA Interoperability with other APIs ​

3.5.4.2 Interprocess Communication ​

3.5.5 Fine-Grained Control ​

3.5.5.1 Virtual Memory Management ​

3.5.5.2 Compute Fabric Transport ​

3.5.5.3 Driver Entry Point Access ​

3.5.5.4 Error Log Management ​