pre copy pinned data to gpu

PaddlePaddle / PaddleDetection

Object Detection toolkit based on PaddlePaddle. It supports object detection, instance segmentation, multiple object tracking and real-time multi-person keypoint detection.

Apache License 2.0

12.35k stars 2.84k forks source link

pre copy pinned data to gpu #8959

Closed wanghuancoder closed 1 month ago

wanghuancoder commented 1 month ago

DataLoader产生的Tensor有很多pinned memory的，在模型训练过程中，每使用一次Dataloader的Tensor就会发生一次同步H2D拷贝，造成CPU打断，进而影响性能。本PR集中、异步将pinned Tensor转到GPU上。

paddle-bot[bot] commented 1 month ago

Thanks for your contribution!