ymd-h / vulkpy

GPGPU array on Vulkan

vulkpy: GPGPU array on Vulkan

vulkpy is a Python package providing GPGPU computation based on Vulkan.

Requirements

• C++20 compatible compiler
• libvulkan
• Vulkan SDK
  • Headers (vulkan/vulkan.hpp and so on)
  • Shaderc (glslc)

On Ubuntu 22.0,

wget -qO - http://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add -
wget -qO /etc/apt/sources.list.d/lunarg-vulkan-focal.list http://packages.lunarg.com/vulkan/lunarg-vulkan-focal.list
apt update
apt install -y libvulkan1 libvulkan-dev vulkan-headers shaderc vulkan-validationlayers

Note
vulkan-sdk cannot be installed because it requires obsolete package qt5-default.

Example

import vulkpy as vk

gpu = vk.GPU()

a = vk.Array(gpu, data=[10, 10, 10])
b = vk.Array(gpu, data=[5, 5, 5])

c = a + b
c.wait()

print(c)
# [15, 15, 15]

Features

• Element-wise Arithmetic Operators between 2 Arrays.
  • +, -, *, /, **, +=, -=, *=, /=, **=
• Arithmetic Operators between Array and float.
  • +, -, *, /, **, +=, -=, *=, /=, **=
• Arithmetic Operators between float and Array.
  • +, -, *, /, **
• Matrix Multiplication Operator between 1d/2d Arrays.
  • @
• Element-wise math functions as Array's member function
  • max(other, inplace=False), min(other, inplace=False)
  • abs(inplace=False), sign(inplace=False)
  • sin(inplace=False), cos(inplace=False), tan(inplace=False)
  • asin(inplace=False), acos(inplace=False), atan(inplace=False)
  • sinh(inplace=False), cosh(inplace=False), tanh(inplace=False)
  • asinh(inplace=False), acosh(inplace=False), atanh(inplace=False)
  • exp(inplace=False), log(inplace=False)
  • exp2(inplace=False), log2(inplace=False)
  • sqrt(inplace=False), invsqrt(inplace=False)
  • clamp(min, max, inplace=False)
• Reduction as Array's member function
  • sum(axis=None), prod(axis=None)
  • maximum(axis=None), minimum(axis=None)
  • mean(axis=None)
  • argmax, argmin
  • ...
• Other Array method
  • gather(idx: U32Array) -> Array
  • tensordot, shuffle
  • ...
• Bloadcast
  • Explicit broadcast copy (memory inefficient, fallback option)
    • broadcast_to(shape) (used at clamp)
  • Special implementations for element-wise arithmetic operators
    • +, -, *, /, **, +=, -=, *=, /=, **=
  • Reduction with re-broadcast
    • sum, prod, maximum, minimum, mean
• Pseudo Random Number Generator (PRNG)
  • xoshiro128++ (vulkpy.random.Xoshiro128pp(gpu, *, size=None, data=None))
    • [0, 1) uniform (.random(shape=None, buffer=None))
    • Gaussian with Box-Muller (.normal(shape=None, buffer=None, mean=0.0, stddev=1.0))
  • pcg32
• Neural Network
  • Layers
    • Dense, ReLU, Sigmoid, Softmax
    • conv, batch norm, layer norm, ...
  • Optimizers
    • SGD, Adam, AdaGrad
    • rmsprop, ...
  • Losses
    • CrossEntropyLoss, SoftmaxCrossEntropyLoss, MSELoss, HuberLoss
    • ...
  • Initializers
    • Constant, HeNormal
    • ...
  • Models
    • Sequance
    • ...
  • Regularization
    • Lasso(coeff=1.0), Ridge(coeff=1.0), Elastic(L1=1.0, L2=1.0)
  • ONNX support
  • Custom user layer with automatic backward() definition.
  • Define by Run API