Skip to contents

High-performance Computing in GW modelling

Source: vignettes/other-01-hpc.Rmd
other-01-hpc.Rmd

Geographically weighted modelling methods can be time consuming, especially when the data size is quite large. To solve this problem, this package provided some techniques to speed up the computing:

  • Multithreading based on OpenMP
  • GPU comoputing based on NVIDIA CUDA computing toolkit

Currently, almost all algorithms have a serial implementation and one or two high-performance implementation.

Algorithm Multithreading GPU computing
Basic GWR o o
Multiscale GWR o
GWDR o

All the parallelizable algorithms take two parameters to control parallelization: parallel_method and parallel_arg. The parallel_method is used to pick a high-performance implementation. The valid options are "omp" and "cuda", representing multithreading and GPU computing respectively. The parallel_arg is used to pass arguments. Different parallel methods needs different arguments.

However, not all platform support both techniques, and some platform may have limited performance. When users choosed unsupported method, it will fall back to the serial implementation. Please refer to the table below for additional information regarding platform support.

Platform Multithreading GPU computing
Windows o o*
macOS (Apple Sillicon)
Linux o o

*On Windows, GPU computing is supported through an additional library called “GWmodelCUDA”, which is built with MSVC. This is because CUDA does not support the MinGW compiler, which is used to compile R packages. To enable GPU computing, please mannually build the package following the instructions bellow.

Mutlithreading

Multithreading implementation is automatically enabled in package on CRAN. To use multithreading implementation, users need to specify number of threads through parallel_arg. The number of threads needs to be determined according to power of the CPU.

gwr_basic(
  formula = PURCHASE ~ FLOORSZ + UNEMPLOY + PROF,
  data = LondonHP, bw = 64, adaptive = TRUE,
  parallel_method = "omp", parallel_arg = 8
)

GPU Computing

Usage

The usage of GPU computing is quite similar to that of multithreading. Users needs to set parallel_method="cuda" and pass a integer vector of 2 elements through parallel_arg. The first element indicates which GPU is used (index starting from 0), and the second element — group size — indicates how many samples are solved together. The group size needs to be selected according to the memory size of the selected GPU. A larger group size will consume more GPU memory. However, experience shows that this number only has a slight effect on efficiency. So set it to 64 will be apporiate for most GPUs.

gwr_basic(
  formula = PURCHASE ~ FLOORSZ + UNEMPLOY + PROF,
  data = LondonHP, bw = 64, adaptive = TRUE,
  parallel_method = "cuda", parallel_arg = c(0, 64)
)

Setup

To enable GPU computing, users need to setup the package mannually.

For Windows

  1. Download the source codes of this package from GitHub.
  2. Download the pre-compiled gwmodelcuda.dll with the bundled dynamic loaded libraries and put them into the src folder within the package’s source code.
  3. Add an environment ENABLE_CUDA with value 1.

  4. Run the following code in command promopt or powershell to build and install the package.

    R.exe CMD INSTALL GWmodel3

If the installation is unsuccessful, please make sure that you are using the lastest R and Rtools, or mannually install the GSL library.

The gwmodelcuda.dll is able to be built by yourself following the steps bellow.

  1. Install CMake and CUDA.
  2. Get dependencies: armadillo, GSl, openblas, catch2 (via conda, vcpkg, and any other ways).
  3. Download the source codes of libgwmodel

  4. Make a new directory build (or any name as you like) and configure the project with the following command within the new directory

    cmake .. \
  -DENALBE_CUDA=ON \
  -DUSE_CUDA_SHARED=ON \
  -DCMAKE_CUDA_ARCHITECTURES=75 \
  -DWITH_TESTS=OFF
    Note the value of CMAKE_CUDA_ARCHITECTURES varies on different GPUs. Please find the official document for more information.

  5. Build the library

    cmake --build . --config Release --target gwmodelcuda

  6. Copy the library file at src/Release/gwmodelcuda.lib and dynamic load libraries under bin/Release to the src directory of GWmodel3.

For Linux

It would be quite easier to setup CUDA on Linux. Install the package withe the following configure argument: --enable-cuda=yes. After getting source package of GWmodel3, users can install this package with the fowllowing commands.

Or install via the remotes package:

remotes::install_github("GWmodel-Lab/GWmodel3", build_opts = c("--configure-args=--enable-cuda=yes"))

Before installing the package, make sure that the environment variable CUDA_HOME is defined as the path to the CUDA toolkit.