PyTorch-Extensions-for-GPU-CPU

This repo contains PyTorch Extensions for 2D convolutional operations written with C++, OpenGL, OpenMP, and Boost Compute. The motivation behind writting faster extensions is that that currently PyTorch native 2D convolution is implmented using CuDNN,but does not support integrated GPU support currently. Thus I have written extensions for running convolution operations on CPU, integrated GPU and shared CPU/GPU.

Hardware: Intel(R) Core(TM) i7-14700HX (Raptor Lake), Intel(R) UHD Graphics

Test size: 4096x4096 tensor Input, 256x356 tensor Filter

1. iGPU [OpenCL]:

  Native PyTorch runtime: 0.30904 secs
  Extension runtime: 0.05667 secs
  Final Speedup: ~88%

2. CPU [OpenMP] :

  Native PyTorch runtime: 0.30904 secs
  Extension runtime: 0.6036 secs 
  Final Speedup: No increase

2. Shared iGPU-CPU [3:1] [OpenMP/OpenCL] :

  Native PyTorch runtime: 0.30904 secs
  Extension runtime: 0.28835 secs 
  Final Speedup: ~13%