README.md

GNP: Fast and Scalable Tensor Engine

• GNP (GPU Numpy) is a fast and scalable tensor computing library that utilize the parallel computation of GPUs. The library is itself built upon numpy library. numpy is a computing library that runs on CPU only, and GNP boost the performance of this framework by optimizing computation of numpy ndarray on GPU.

• GNP will have its own array structure which is based upon numpy. GNP implements GPU operations by binding python numpy data and use Cuda kernel to parallelize computation on GPU. The list of supported kernel is listed in the later section.

Requirements and Installation:

• Requirements:

  • OS: Linux
  • Python version: 3.10
  • Anaconda

• Installation:

  # First, create a conda environment and install:
  conda create -n gnp pip python=3.10
  conda activate gnp
  # install required libraries
  conda env update -f dev_env.yml
  conda install cudatoolkit
  # Instlal current library
  python setup.py install
  

Project Structure and Supported APIs

Supported Cuda Kernel

• Unary ops:

  • Negative
  • Positive
  • Invert (~x = 1/x)

• Binary ops:

  • add/subtract
  • multiply/true divide/floor divide
  • mod
  • pow
  • and
  • or
  • xor
  • right shift
  • left shift

• Comparison ops:

  • greater than
  • less than
  • greater than or equal to
  • less than or equal to
  • equal
  • not equal

• Linear algebra:

  • Matrix multiplication

Supported APIs

• Experimental:
  • Neural network sequential API
  • Fully connected neural network forward and backward pass
  • SGD Optimizer, MSE Loss
  • Non-linear activations: Relu, Tanh, Sigmoid

Result

• show code
• Compare time
• Future work

Documentations

• The slide is here: https://docs.google.com/presentation/d/1AYxhXKqORdddHHGxe6CHCLwE8b6S0u4CGl_jjvH4tIY/edit?usp=sharing

  • The slide is also in the pdf file in the docs folder

• For the class diagrams, install Visual Paradigm and run the GNP.vpp file in the docs/vpp/ folder