GeForce RTX 4060 is not supported

[EmanH]

I've followed the docs for getting gpu accelerated nvenc encoding, but stuck with:

The RTX 4060 definitely supports encoding.

Debug logs are:

CUDA 12.3.0 / PyCUDA 2024.1, found 1 devices:^[[0m
  + NVIDIA GeForce RTX 4060 @ 0000:01:00.0 (memory: 87% free, compute: 8.9)^[[0m
^[[32mnvmlSystemGetDriverVersion=552.22^[[0m
NVidia driver version 552.22^[[0m
^[[32minit_module() will try keys: [None]^[[0m
^[[32mtesting encoder with device 0^[[0m
^[[32mload_device(0)^[[0m
^[[32mcuda_device_context(0)^[[0m
^[[32mcuda context allocation took 56ms^[[0m
^[[32minit_cuda(<pycuda._driver.Context object at 0x7f5b356ebe60>) pixel format=None^[[0m
^[[32minit_cuda(<pycuda._driver.Context object at 0x7f5b356ebe60>)^[[0m
^[[32minit_cuda cuda info={'driver': {'version': (12, 3, 0), 'driver_version': 12040}}^[[0m
^[[32mtest encoder=nvenc(None/None/None - None -    0x0   )^[[0m
^[[32mopen_encode_session() cuda_context=None, cuda_context_ptr=0x55b571a27fc0^[[0m                                                                                                                              ^[[32mopen_encode_session() functionList=0x55b570d34350^[[0m
^[[32mfailed to test encoder with cuda_device_context(0 - locked)
Traceback (most recent call last):
  File "xpra/codecs/nvidia/nvenc/encoder.pyx", line 3086, in xpra.codecs.nvidia.nvenc.encoder.init_module
  File "xpra/codecs/nvidia/nvenc/encoder.pyx", line 2989, in xpra.codecs.nvidia.nvenc.encoder.Encoder.open_encode_session
AssertionError: looks like NvEncodeAPICreateInstance failed!^[[0m
^[[33m device NVIDIA GeForce RTX 4060 is not supported: looks like NvEncodeAPICreateInstance failed!^[[0m
^[[32mclean() cuda_context=None, encoder context=0x0^[[0m
^[[32mclean() done^[[0m                                                                                                                                                                                          ^[[33mno valid NVENC devices found^[[0m
^[[32mxpra.codecs.nvidia.nvenc.encoder.log=Logger['xpra.codecs.loader', 'encoder', 'nvenc']^[[0m
^[[32mxpra.codecs.nvidia.nvenc.encoder.Encoder=<class 'xpra.codecs.nvidia.nvenc.encoder.Encoder'>^[[0m                                                                                                           ^[[32mnvenc.selftest=<cyfunction selftest at 0x7f5b3579f9f0>^[[0m
^[[32mhas_nvidia_hardware()=None^[[0m
^[[32m<cyfunction selftest at 0x7f5b3579f9f0> failed                                                                                                                                                             Traceback (most recent call last):
  File "/usr/lib64/python3.12/site-packages/xpra/codecs/loader.py", line 147, in codec_import_check
    selftest(FULL_SELFTEST)                                                                                                                                                                                        File "xpra/codecs/nvidia/nvenc/encoder.pyx", line 3226, in xpra.codecs.nvidia.nvenc.encoder.selftest
ImportError: no nvidia GPU device found^

System is Fedora 40.

The encode SDK I'm using has the following requirements, which as far as I can see are being met.

[EmanH]

This is using the pre-packaged xpra packages.

I've also tried building from source with nvenc on Ubuntu 22.04.

It gets to the same place and fails with device NVIDIA GeForce RTX 4060 is not supported: looks like NvEncodeAPICreateInstance failed!

I'm running inside a docker container.

[totaam]

Try newer / older drivers and test without using a docker container, this should narrow it down a bit.

[EmanH]

@totaam I made some progress with the following docker image:

FROM nvcr.io/nvidia/cuda:12.4.1-runtime-ubuntu22.04

ENV NVIDIA_VISIBLE_DEVICES all
ENV NVIDIA_DRIVER_CAPABILITIES video,compute,utility

# Install dependencies
RUN apt-get update && apt-get install -y \
    python3-pip \
    && rm -rf /var/lib/apt/lists/*

RUN apt-get update && apt-get install -y cuda-compiler-12-4 cuda-profiler-api-12-4 libcurand-dev-12-4

# # Install pycuda
RUN pip3 install pycuda

RUN apt-get install -y wget gpg

RUN apt-get update && \
    apt-get install -y software-properties-common && \
    add-apt-repository universe && \
    add-apt-repository multiverse && \
    apt-get update

RUN mkdir -p /usr/lib/pkgconfig
RUN apt-get install -y pkg-config

RUN apt-get install -y xvfb python3-cairo python3-gi-cairo python3-opengl python3-pil
RUN apt-get install -y libx11-dev libxtst-dev libxcomposite-dev libxdamage-dev libxres-dev \
      libxkbfile-dev \
      python3-dev \
      pandoc \
      libsystemd-dev \
      liblz4-dev
RUN apt-get install -y libgtk-3-dev python3-cairo-dev python-gi-dev
RUN apt-get install -y libxxhash-dev
RUN pip install Cython pynvml numpy
ENV XPRA_NVENC_DEBUG_API=1
RUN wget https://github.com/Xpra-org/xpra/archive/refs/tags/v6.0.1.tar.gz && \
  tar -xvzf v6.0.1.tar.gz && \
  rm v6.0.1.tar.gz

COPY ./nvenc /opt/nvenc
COPY ./nvenc.pc /usr/lib/pkgconfig/nvenc.pc
COPY ./build_cuda_kernels.py ./xpra-6.0.1/fs/bin/build_cuda_kernels.py

RUN cd xpra-6.0.1 && ./setup.py build --with-nvenc && ./setup.py install

It now initializes NVENC successfully:

But it doesn't list h264 or h265 (nvenc) as supported server options:

xpra start --encoding=help
Warning: using '/run/user/1000' as XDG_RUNTIME_DIR
xpra server supports the following encodings:
(please wait, encoder initialization may take a few seconds)
2024-06-05 17:43:01,404 Warning: failed to import GStreamer 1.x:
2024-06-05 17:43:01,404  Namespace Gst not available
 * grayscale   same as 'auto' but in grayscale mode
 * png         Portable Network Graphics (lossless, 24bpp or 32bpp for transparency)
 * png/P       Portable Network Graphics (lossy, 8bpp colour)
 * png/L       Portable Network Graphics (lossy, 8bpp grayscale)
 * rgb         Raw RGB pixels, lossless , may be compressed using lz4 (24bpp or 32bpp for transparency)
 * jpeg        JPEG lossy compression
 * scroll      motion vectors, supplemented with picture codecs

We're running xpra with a command like:

xpra start -d damage,compress,encoding :$displayNumber --start="bash ./run.sh" --bind-ws=0.0.0.0:$portNumber --video-encoders=nvenc --encoding=h264 --daemon=no --sharing=no --start-new-commands=no 2>&1

It seems to say nvenc is happy, but it doesn’t list it as an available module:

When debugging encoding, I can see it is using rgb24, rgb32 and jpeg via the native pillow encoder.

add_encodings('rgb24', 'rgb32')
...
add_encoding(jpeg)
...
pillow encodings: ('png', 'png/L', 'png/P', 'jpeg', 'webp')

So maybe I'm missing something, or not building something correctly?

I do run ./setup.py build --with-nvenc and it doesn't complain.

[totaam]

Looking at your long log sample, I think it is getting tripped by the has_nvidia_hardware checks.
But you can easily bypass that:

xpra/xpra/codecs/nvidia/util.py

Line 23 in 683a8ec

NVIDIA_HARDWARE = envbool("XPRA_NVIDIA_HARDWARE", False)

with:

$ XPRA_NVIDIA_HARDWARE=1 XPRA_ALL_DEBUG=1 xpra encoding nvenc

[EmanH]

nvidia-smi might be a good way to determine the presence of nvidia hardware in containers running on WSL2 (and other contexts I think). It's shipped with the nvidia container toolkit

This works:

    try:
        output = subprocess.check_output(["nvidia-smi", "--query-gpu=name", "--format=csv,noheader"])
        output = output.decode("utf-8").strip()
        if output:
            log(f"has_nvidia_hardware() found NVIDIA GPU(s) using nvidia-smi: {output}")
            return True
    except (FileNotFoundError, subprocess.CalledProcessError):
        log("has_nvidia_hardware() nvidia-smi command not found or failed")

I might raise a PR anyhow, and see what you think.

[EmanH]

Yeah, with XPRA_NVIDIA_HARDWARE=1 it shows the nvenc module no problem.

XPRA_NVIDIA_HARDWARE=1 xpra encoding nvenc
Warning: using '/run/user/1000' as XDG_RUNTIME_DIR
CUDA initialization (this may take a few seconds)
CUDA 12.4.0 / PyCUDA 2024.1, found 1 devices:
  + NVIDIA GeForce RTX 4060 @ 0000:01:00.0 (memory: 87% free, compute: 8.9)
NVidia driver version 552.22
NVENC v12 successfully initialized with codecs: h264, h265
modules found:
* nvenc                : /usr/local/lib/python3.10/dist-packages/xpra/codecs/nvidia/nvenc/encoder.cpython-310-x86_64-linux-gnu.so

codecs versions:
* nvenc                           : 12.2

But it's still using xpra.codecs.pillow.encoder when starting xpra:

XPRA_NVIDIA_HARDWARE=1 xpra start -d damage,compress,encoding :$displayNumber --start="bash ./run.sh" --bind-ws=0.0.0.0:$portNumber --video-encoders=nvenc --encoding=h264 --daemon=no --sharing=no --start-new-commands=no 2>&1

enc_pillow=<module 'xpra.codecs.pillow.encoder' from '/usr/local/lib/python3.10/dist-packages/xpra/codecs/pillow/encoder.py'>
pillow encodings: ('png', 'png/L', 'png/P', 'jpeg', 'webp')

[totaam]

But it's still using xpra.codecs.pillow.encoder when starting xpra:

Yes, that's correct.
It will only use nvenc when needed, which is why you need the other encoders to fill the gaps.

[totaam]

I might raise a PR anyhow, and see what you think.

Here it is:
#4257

[totaam]

Oh, and BTW, since you're using containers, are there any specific reasons for using Debian or Ubuntu?
These are the worst distributions for running xpra.

[EmanH]

The cuda base images are only available for ubuntu, red hat universal or rocky linux. I don't think we can use red hat and don't know anything about Rocky Linux. It does make it easier using one of these base images than building from scratch.

[totaam]

rockylinux would work better than any Debian based distro - they don't butcher Xorg so Xdummy will work without any hassle.
Fedora also has the cuda bits in rpmfusion.

[EmanH]

Yeah ok, maybe we'll try rockylinux

[Simba98]

Hi EmanH,

I have some issues on running nvenc on RockyLinux 8 for compile xpra.
If you have tries rockylinux, could you share your dockerfile and / or nvenc bin or scripts. Thank you very much!

[EmanH]

@Simba98 We didn't end up trying RockyLinux sorry