Skip to content

INSTALL_CMDSTAN.md

Latest commit

 

History

History
103 lines (72 loc) · 4.87 KB

INSTALL_CMDSTAN.md

File metadata and controls

103 lines (72 loc) · 4.87 KB
Raw

Installing CmdStan with GPU support

These instructions assume that you have an NVIDIA or AMD GPU in your system. While modern Intel GPUs support OpenCL, their double-precision performance is unlikely to bring speedup compared to a modern CPU.

STEP 1: Install the toolchain and GPU driver

Ubuntu

Make sure you have make, g++ 4.9.3 or higher (or clang++ 6.0 or higher). If using a NVIDIA device, install the NVIDIA CUDA toolkit and clinfo tool.

apt update
apt install nvidia-drivers-460 nvidia-cuda-toolkit clinfo

Replace the driver version (460 in the above case) with the latest version.

If you have an AMD GPU, install the Radeon Open Computer platform (ROCm). Instructions can be found here.

Windows

Make sure you have mingw32-make and g++ 4.9.3 or higher. If not, you can install both as part of the Rtools 4.0 suite. During installation make sure that the 64-bit toolchain is installed. Upon installing RTools 4.0, open the RTools bash and enter pacman -Sy mingw-w64-x86_64-make. This will install mingw32-make. Add C:\RTools40\usr\bin and C:\RTools40\mingw64\bin to the PATH environment variable if they are not added automatically.

In order to make sure both mingw32-make and g++ are installed, try out the mingw32-make --version and g++ --version commands.

If you have an NVIDIA device, install the latest NVIDIA CUDA toolkit found here. AMD users can use the OCL-SDK.

STEP 2: Determine the IDs of your target device

In order to run an OpenCL application, you need to specify the platform and device IDs of your target device. If you only have a single OpenCL-enabled device (most likely a GPU) both IDs are 0. In this case you can proceed to Step 3.

If you have installed OpenCL runtimes (drivers) for multiple GPUs or have previously installed the OpenCL runtime for your CPU, you first need to determine the platform and device ID of your target device. You can use the clinfo tool to list all OpenCL-enabled devices and their IDs. On Linux clinfo can be obtained using sudo apt install clinfo or equivalently using yum. You can also build the tool from the source which is available at here.

Windows binaries of clinfo are available here.

STEP 3: Install and configure CmdStan

Download the latest release of CmdStan from here. Download and untar the .tar.gz file.

Create a file named local inside the make folder and populate it with the following:

STAN_OPENCL=true
OPENCL_PLATFORM_ID=#PLATFORM_ID
OPENCL_DEVICE_ID=#DEVICE_ID

where you replace #PLATFORM_ID and #DEVICE_ID with the IDs from step 2.

On Windows you also need to specify the path to the OpenCL.lib file by adding the following (replace $(PATH) with the actual path to the file):

LDFLAGS_OPENCL= -L$(PATH) -lOpenCL

If you are using and AMD GPU with the OCL SDK suggested in step 1 the LDFLAGS_OPENCL will typically be:

LDFLAGS_OPENCL= -L"C:/Program Files (x86)/OCL_SDK_Light/lib/x86_64" -lOpenCL

If using and NVIDIA GPU then a typical LDFLAGS_OPENCL will be (depending on the installed version):

LDFLAGS_OPENCL= -L"C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v10.2/lib/x64" -lOpenCL

STEP 4: Build CmdStan

On Linux run make build. On Windows run mingw32-make build. If you have multiple cores available you can also use the -j flag to specify building with multiple cores. Example make build -j 4 or mingw32-make build -j 4 will use 4 cores to build cmdstan.

STEP 5: Compile and run the logistic regression model example

Linux

First generate the input data for the logistic regression model. You can do that by running

Rscript generate_glm_data.R

from the lr_glm folder. The script requires the jsonlite package that can be installed by running install.packages("jsonlite").

If you place the lr_glm folder in the home (~) folder compile from the cmdstan folder with

make ~/lr_glm/lr_glm

and start sampling with

./lr_glm sample num_samples=500 num_warmup=500 random seed=1 data file=lr_glm.data.json
Windows

First generate the input data for the logistic regression model. You can do that by running

Rscript.exe generate_glm_data.R

in the lr_glm folder. The script requires the jsonlite package that can be installed by running install.packages("jsonlite").

If you place the lr_glm folder in the home (~) folder compile from the cmdstan folder

mingw32-make ~/lr_glm/lr_glm.exe

and start sampling with:

~/lr_glm/lr_glm.exe sample num_samples=500 num_warmup=500 random seed=1 data file=~/lr_glm/lr_glm.data.json