Issue with GStreamer Python plugin accelerated with PyCUDA

Accelerated Computing CUDA CUDA Programming and Performance
,
1

Hi,

I am trying to create a GStreamer plugin using the python bindings that would use PyCUDA to accelerate processing but the PyCUDA calls will fails inside some of the virtual methods. This is weird because I have no problems calling GPU-accelerated PyTorch.

My setup:

  • Jetson Nano with JetPack 4.6
  • Over a clean JetPack I intalled:
sudo apt install \
python-gi-dev \
python-gst-1.0 \
gstreamer1.0-python3-plugin-loader \
gstreamer1.0-plugins-base \
libgstreamer1.0-dev \
python3-pip

pip3 install cython
pip3 install numpy pycuda

Here is a minimal example that reproduces the issue:

import gi
gi.require_version('Gst', '1.0')
gi.require_version('GstBase', '1.0')
from gi.repository import Gst, GObject, GstBase
from pycuda.compiler import SourceModule
import numpy
import pycuda.autoinit
import pycuda.driver as drv

Gst.init(None)

class GstCUDAAdd(GstBase.BaseTransform):
    """
    Class that inherits from BaseTransform
    """

    __gstmetadata__=(
        'CUDAAdd',
        'Generic',
        'Add using PyCUDA',
        'Miguel Taylor <miguel.taylor@ridgerun.com>')

    __gsttemplates__=(
        Gst.PadTemplate.new(
            "src",
            Gst.PadDirection.SRC,
            Gst.PadPresence.ALWAYS,
            Gst.Caps.new_any()),
        Gst.PadTemplate.new(
            "sink",
            Gst.PadDirection.SINK,
            Gst.PadPresence.ALWAYS,
            Gst.Caps.new_any()))

    __gproperties__={}

    def __init__(self):
        """
        Initialize new GstCUDAAdd
        """
        self.add()
        GstBase.BaseTransform.__init__(self)
        GstBase.BaseTransform.set_in_place(self, False)

    def do_transform(self, inbuf, outbuf):
        """
        Implementation of GstBaseTransform class 'transform'

        Parameters
        ----------
        inbuf: Gst.Buffer
            Input buffer (read only)
        outbuf: Gst.Buffer
            Output buffer (read only, allocated)

        Returns
        ----------
        return: Gst.FlowReturn
            The result of passing data to a pad.
        """
        self.add()
        return Gst.FlowReturn.OK


    def add(self):
        # translate function to python 
        mod = SourceModule("""
		__global__ void add_them(float *dest, float *a, float *b)
		{
			const int i = threadIdx.x;
			dest[i] = a[i] + b[i];
		}
		""")
        add_them = mod.get_function("add_them")

        # make 2 sets of 400 random floats
        a = numpy.random.randn(400).astype(numpy.float32)
        b = numpy.random.randn(400).astype(numpy.float32)

        # create a set of 0s
        dest = numpy.zeros_like(a)

        # replace 0s with results of a + b
        add_them(drv.Out(dest), drv.In(a), drv.In(b), block=(400,1,1), grid=(1,1))

        # should print block of 0s -> (a+b) - (a+b)
        print(dest - (a + b))


GObject.type_register(GstCUDAAdd)
__gstelementfactory__=("cudaadd", Gst.Rank.NONE, GstCUDAAdd)

If I call self.add() in the class init it works correctly, but inside other methods like do_transform or do_prepare_output_buffer it will fail with one of the following errors:

  File "/home/nvidia/.local/lib/python3.6/site-packages/pycuda/compiler.py", line 363, in __init__
    self.module = module_from_buffer(cubin)
pycuda._driver.LogicError

  File "/home/nvidia/.local/lib/python3.6/site-packages/pycuda/compiler.py", line 363, in __init__
    self.module = module_from_buffer(cubin)
pycuda._driver.LogicError: cuModuleLoadDataEx failed: invalid device context -

Do you know what could be going on or how to fix this issue?

2

In my experience this is usually a problem with pycuda.autoinit that creates a CUDA context separate from the one you are trying to use (elsewhere). I don’t know for sure that this is the issue here, I have not tried your code. See here.

2 Likes
3

Thanks @Robert_Crovella,

Changing the CUDA init and using retain_primary_context solved the issue:

cuda.init()
pycuda_ctx = cuda.Device(0).retain_primary_context()
pycuda_ctx.push()
...
pycuda_ctx.pop()
4

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.