Skip to content

Latest commit

 

History

History
92 lines (60 loc) · 4.37 KB

README.md

File metadata and controls

92 lines (60 loc) · 4.37 KB
Raw

Pytorch OpenCL backend based on dlprimitives

DLPrimitives-OpenCL out of tree backend for pytorch

It is only beginning, but you can train some vision nets using OpenCL devices.

Supported pytorch versions are 1.13 and torch >= 2.4

Validated Networks

Following torchvision networks were validated:

Network Notes
alexnet
resnet18
resnet50
convnext_small
vgg16
squeezenet1_0
googlenet
densenet161
inception_v3 fwd only - backward fails on cuda/cpu
shufflenet_v2_x1_0
mobilenet_v2
mobilenet_v3_large
mobilenet_v3_small fwd only - same failure on bwd on cuda.
resnext50_32x4d
wide_resnet50_2
mnasnet1_0
efficientnet_b0
efficientnet_b4
regnet_y_400mf
vit_b_16

Calculations validated agaist CPU reference for both forward and backward popogation.

Tested Devices

DLPrimitves itself is tested on following devies:

  • AMD rx 6600XT with ROCM drivers, rx560 16cu with AMDGPU-pro drivers
  • Nvidia: GTX 960
  • Intel: HD 530, UHD 770, Arc A380

Installation

The simplest way is to download whl package from release page - builds for Windows and Linux are provided.

Create virtual evironment, install CPU version of pytorch >= 2.4, download whl file that corresponds to python version, pytorch version and architecture and install it. For example: python 3.10, torch 2.4 on Linux it is:

pip install pytorch_ocl-0.1.0+torch2.4-cp310-none-linux_x86_64.whl

Now all you need is to import pytorch_ocl and now you can use device 'ocl' instead of cuda:

torch.randn(10,10,device='ocl:0')

If you don't have prebuilt whl file for your enviromnent, you want use latest updates or modify the code, please refer to README-build.md

How to Use

Keep it mind... it is earky version that misses a lot of functionality and it isn't fully tested yet. So if something fails. It is either not implemented or it is implemented incorrectly

Note: pytorch backend is based on dlprimitives library that actually implements all the operators and it is relatively well tested.

If you still want to try: import package pytorch_ocl

Keep in mind you may have several OpenCL devices. Refer to clinfo --list to list of the devices and their order. Now instead of calling something.to('cuda') you call something.to('ocl:0') or something.to('privateuseone:0' for pytorch 1.13) or another ocl:1 etc.

Known Issues

  1. Many operators not implemented and there may be fallbacks to CPU. Sometimes it is minor but sometimes it may hamper the performance, some may just fail
  2. When you save/restore the model move it to CPU. Currently there is an issue with loading back saved state dictionary if it was saved from ocl device

pytorch_ocl specific API

Some functions specific to pytorch_ocl. When using pytorch >= 2.4 they are accessible from torch.ocl and pytorch_ocl, for 1.13 you must use pytorch_ocl

  • torch.ocl.empty_cache(): Same as torch.cuda.empty_cache() remove all cached GPU memory
  • torch.ocl.synchronize(device=None): synchronize all operations queue on the device, if device is None - all of them same as torch.cuda.synchonize
  • torch.ocl.manual_seed_all(seed): reset random number generator state. torch.manual_seed - it calls automatically for pytorch >= 2.4. Note for pytorch 1.13 you must call pytorch_ocl.manual_seed_all