DSCOptimization

Implementation of Optimized Depthwise Separable Convolution Kernels and PyTorch Extensions on Deep Computing Unit

Introduction

Depthwise Separable Convolution decomposes the standard convolution operation into “Depthwise Convolution” and “Pointwise Convolution” to reduce computational overhead. This method has been applied to various deep learning neural network models, such as the MobileNet series and EfficientNet series. This project focuses on optimizing the depthwise separable convolution. The optimized convolution kernel functions are encapsulated as PyTorch extensions, allowing them to be called within PyTorch.

Directories

• Depthwise

  • Kernel: kernels and tests for depthwise convolution
  • Extension
    • DCU_Depthwise_Extension: pytorch extensions for depthwise convolution
    • Test_DCU_Depthwise_Extension: tests for depthwise extensions

• Pointwise

  • Kernel: kernels and tests for pointwise convolution
  • Extension
    • DCU_Pointwise_Extension: pytorch extensions for pointwise convolution
    • Test_DCU_Pointwise_Extension: tests for pointwise extensions
    • AutomaticPointwiseHardwareSelector: Tiling Parameter Generator + Code Generator + Profiler