CUDA Graph 在 TensorFlow 中的应用.pdf

《CUDA Graph 在 TensorFlow 中的应用.pdf》由会员分享，可在线阅读，更多相关《CUDA Graph 在 TensorFlow 中的应用.pdf（28页珍藏版）》请在三个皮匠报告上搜索。

1、#page#CONTENTWhats CUDA GraphHow to Use CUDA GraphLaunch Overhead in TensorFlowIntegrate CUDA Graph into TensorFlowPerformance#page#WHATS CUDA GRAPHWhat Problem CUDA Graph SolvesReduce Launch OverheadsDBgraph launchBD#page#WHATS CUDA GRAPHStream Launch vs Graph LaunchStream LaunchStreamQueuesExecuti

2、onGrfdManagementCUDAA）BlockAO0WSs22BlockA1Dxx6,s2(）SM1GrCompetionhttps:/dem/gtc/2020/video/s21760#page#WHATS CUDA GRAPHStream Launch vs Graph LaunchGraph Launch (Pre-Ampere）CUDA Graph LaunchStreamQucuesGridManagementExecutionBlockAOSMOBlockA1SM1cudaGraphLuanchg1，51）mOther DependenciesGrid Completion

3、https:/dem/gtc/2020/ido/21760#page#page#WHATS CUDA GRAPHStream Launch vs Graph LaunchLaunch overhead comparison (test using empty kernel）A100 GPU *Graph with 32 nodesgraphstreamPatternhost(ms） device (ms）host(ms）device (ms）host speedupdevice speedup4.4314.72.21striaght line28.1265.2560.672two branch

4、es21.85.43.1715.4769.2583.4693.7521.97.63fork andjoin4.2821.32161.79#page#HOW TO USE CUDA GRAPH口 Define a CUDA Graph口 Stream Capture口 CUDA Graph API口Instantiate a CUDA Graph口Call cudaGraphlnstantiate(）口Launch the CUDA Graph executable instance口Call cudaGraphLaunch(.）https:/docCUDARTGRAPH.html#page#H

5、OW TO USE CUDA GRAPHDefine a CUDA GraphStream CaptureGraph APIScudaStreamBeginCapture（stream1）/Createthegraphitstartsstream2emptystreamcudaGrapnCrtcudaEventRecordevent1，stream1nt1）anodeParams）BcudaEventRecordlevent2.stream2）/ Now set up dependencies on each node/A-BcudaGraphAddDependencies(sraph，a，8

6、c，/A-Cdssuudppudeep/B-DcudaGraphAddDependenciesgraph，ac,ad,1）/C-D/End capture in the origin stream#page#HOW TO USE CUDA GRAPHCUDA Graph Node TypesKernelCPU function callMemory copyMemsetEmpty nodeChild graph#page#LAUNCH OVERHEAD IN TENSORFLOWTF op SchedulingNode A is ready to Runready_nodes = A，Call