In order to run the performance tests you need to have a proper toolchain and OpenCL-enabled GPU driver installed.
If using a NVIDIA device install the NVIDIA CUDA toolkit and clinfo tool.
apt update
apt install nvidia-cuda-toolkit clinfo
If you have an AMD GPU install the OpenCL driver available here:
https://www.amd.com/en/support/kb/release-notes/rn-prorad-lin-amdgpupro
Install the latest Rtools suite if you don't already have it. During the installation make sure that the 64-bit toolchain is installed. You also need to verify that you have the System Enviroment variable Path updated to include the path to the g++ compiler (\mingw_64\bin).
If you have an NVIDIA device install the latest NVIDIA CUDA toolkit found on the NVIDIA support website. AMD users should use the AMD APP SDK.
Download the Stan Math 3.1.1 release archive and unzip/untar it. Further instructions label the installation folder of Stan Math as $(STAN_MATH_PATH). Replace the label with the actual path on your system.
Enable OpenCL in Stan Math. Create a file named local in $(STAN_MATH_PATH)/make/. Add the following lines to the created files:
STAN_OPENCL=true
OPENCL_DEVICE_ID=0
OPENCL_PLATFORM_ID=0
In the unlikely case that your system has more than one OpenCL-enabled device use the clinfo tool to determine the actual IDs and replace the zeros in the make/local file. On Linux clinfo can be obtained using sudo apt install clinfo or equivalent. You can also build the tool from source that is available here.
Windows binaries of clinfo are available here.
On Windows you need to add an additional line that links the OpenCL library. First locate the the OpenCL.lib file and then add the following line, where you replace $(OPENCL_LIB_PATH) with the actual path to the OpenCL.lib file:
LDFLAGS_OPENCL= -L"$(OPENCL_LIB_PATH)/" -lOpenCL
Example:
LDFLAGS_OPENCL= -L"C:/Program Files/OCL_SDK_Light/lib/x86_64" -lOpenCL
First build the Math library dependencies using
make -f $(STAN_MATH_PATH)/make/standalone math-libs
Move to the replicationScripts/Cholesky folder and build the performance tests
make -f $(STAN_MATH_PATH)/make/standalone cholesky_prim_perf cholesky_rev_perf cholesky_rev_grad_perf
On Windows use mingw32-make instead of make. mingw32-make is part of RTools for Windows.
Run the performance tests for Eigen by running ./cholesky_prim_perf, performance tests for Eigen by running ./cholesky_rev_perf and for the gradients using by running ./cholesky_rev_grad_perf.