cuda-toolkit | GitHub Action to install CUDA | GPU library

Just simply do a clean install of CUDA (First delete all of its traces and then try to reinstall the correct version, which is 10.1).

OR:- you can install Miniconda or Anaconda on your computer and then run this command on cmd :-

conda create --name tf_gpu tensorflow-gpu

which would automatically download and install all the correct software needed to run TensorFlow on GPU.

The problem likely comes from the very small amount of data and the loop dependency. Indeed, modern Nvidia GPUs can execute thousands of CUDA threads simultaneously (packed in warps of 32 threads) thanks to the large amount of CUDA cores. In your case, each thread performs a computation on one cell of array_out using a sequential loop. However, there are only 256 cells. Thus, at most 256 threads (8 warps) can run simultaneously - only a tiny faction of the number of simultaneous threads that your GPU should be able to manage. As a result, if you want a better speed-up, you need to provide more parallelism to the GPU (for example by increasing the data size or by computing multiple regression at the same time).

This suggestion is an alternative to try without emulator

https://sourceforge.net/projects/toysbox/files/bionic-nvidia/ubuntu-20.04-5.4.0-26-generic-nvidia-450.57-primeselect.iso

it is live iso-image with nvidia-450.57 installed so you can run your cuda directly ;the only requirement is to set PATH and LD_LIBRARY_PATH to make it awared of your cuda runtime path ; in particular don't forget a link so cuda seems to find its compiler at /usr/local/cuda/bin

just use it on usb stick or simply boot from iso image using grub loopback mechanism.

hoan

First of all - I have the experience limited to ubuntu 18.04 and 16.xx and python DL frameworks. But I hope some sugestions will be helpfull.

• If I were familiar with docker I would rather consider to use docker instead of setting-up everything from scratch. This approach is described in section about tensorflow container
• If you decided to setup all components yourself please see this guideline I used some contents from it for 18.04, succesfully.
• be carefull with automatic updates. After the configuration is finished and tested protect it from being overwritten with newest version of CUDAor TensorRT.

Answering one of your sub-questions - How does this integrate with jupyter notebook? - it does not, becuase it is unneccesary. CUDA library cooperates with a framework such as Tensorflow, not with the Jupyter. Jupyter is just an editor and execution controller on the server side.

NVidia Cuda v3.2 was released according to this on Nov 2010 and OpenCL 1.2 spec was released one year later on November 15, 2011. So I suspect cmake is detecting OpenCL 1.1 correctly.

If you have another SDK installed and you want cmake to detect OpenCL 1.2 despite having another SDKs supporting older version you need to specify that information in cmake. Otherwise it will find the first OpenCL on the search path and stop. So it should be specified find_package(OpenCL 1.2 REQUIRED) or as @squareskittles pointed find_package(OpenCL 1.2 EXACT REQUIRED) if you want exact version.

However you may need to add other SDKs paths to PATH or specify them in cmake so that it has a chance to examine other OpenCL versions. If you have a look at find cmake macros content they contain some typical search paths specified and if you have SDK installed in other not standard path you have to tell that cmake yourself. That is especially the case on Windows where you don't have standard more specific install paths for includes or libraries like for example on Linux. On Windows there is really Program Files but that is too generic and cmake would have to search through it recursively and I'm not sure if that is even supported.

I suspect you may have nvidia cuda 3.2 toolkit path added to PATH only or you specified that path in cmake only. So here would the problem lie. Adding other SDKs paths may resolve the issue.

Also I think clinfo checks runtime OpenCL installations meaning it can be any vendor OpenCL.dll which supports OpenCL 1.2 on your NVidia GPU and cmake checks in SDK's header which OpenCL version your installed SDK supports. So here can be the discrepancy. In this case you may need to install newer cuda toolkit.

This seems to be a known bug in tensorflow, it has something to do with memory allocation that tensorflow is doing in 20XX cards. See detailed thread here:

https://github.com/tensorflow/tensorflow/issues/24496

What fixed the problem for me is adding the following code at the top of my script:

From Could not dlopen library 'libcudart.so.10.0'; we can get that you tensorflow package is built against CUDA 10.0. You should install CUDA 10.0 or build it from source (against CUDA 10.1 or 10.2) by yourself.

As pointed out in comments, if this card is not set to TCC mode, then it cannot be used for CUDA debugging in Windows using next generation debugging (which is all the Turing cards support).

My solution is that I added another NVIDIA card to my computer to dedicate on display, so my Quadro 4000 RTX can be focused on computation (TCC mode). It works perfect now.

I think I found the solution/reason

Writing it here in case someone finds themselves in similar situation, plus I hate finding old threads with similar problem and no answer or "nevermind, I solved it" as the only follow up

Docker build does not have any ties to nvidia runtime and gstreamer requires access to the full nvidia toolchain in order to build the plugins that need it. This is to be resolved with gstreamer 1.18 but until then, there is no way to build gstreamer with nvidia codecs in docker build.

The workaround:

1. Build image with all dependencies.
2. Run a container of said image using runtime="nvidia" but don't use --rm flag
3. In the container, build gstreamer and install it as normally.
4. Verify with gst-inspect-1.0
5. Commit the container as new image: docker commit
6. Tag the temporary image properly.