I am trying to run a simple c++ program, with Cuda Thrust functions, on WSL2. It seems that program fails in runtime to allocate device memory. I use Thrust with Microsoft visual studio all the time, and I don’t get any errors.
CMakeLists.txt:
cmake_minimum_required(VERSION 3.8 FATAL_ERROR)
project(proj LANGUAGES CXX CUDA)
add_executable(proj
proj.cu
)
proj.cu:
#include<thrust/host_vector.h>
#include<thrust/device_vector.h>
#include<thrust/fill.h>
int main()
{
//thrust::host_vector<int> h_vec(10);
//thrust::fill(h_vec.begin(), h_vec.end(), 1);
thrust::device_vector<int> d_vec(10);
//thrust::fill(d_vec.begin(), d_vec.end(), 1);
return 1;
}
output:
terminate called after throwing an instance of 'thrust::system::system_error'
what(): get_max_shared_memory_per_block :failed to cudaGetDevice: unknown error
Aborted (core dumped)
If I comment the device_vector line, and use the host vector instead, it runs with no errors.
Additional info:
Before I post the question, I had already seen the instructions in here , and both downloaded and installed the CUDA driver for WSL, and joined the windows insider program and upgraded to windows 11 build. It did not work though.
But I also had Ubuntu 18.04 installed. I think it was installing Ubuntu 20.4 instead that made the program run without errors in the end!
So, this is what I did to solve this:
-installed Ubuntu 20.04 from Microsoft store (It asked me to download and run a WSL fix first, which I did)
-In powershell, I removed Ubuntu 18.04
wsl --unregister Ubuntu-18.04
Replace Ubuntu-18.04 with the name of your distribution. You can get its name in powershell by
wsl --list
-installed cmake and g++-9 (gcc-9 was already installed)
-downloaded the cuda toolkit using the instructions in here
Then
cmake -DCMAKE_CUDA_COMPILER=/usr/local/cuda-11.0/bin/nvcc ..
make
./proj