Please find the configurations that I use.
• x86/orin
• DeepStream 7
• JetPack Version (5.1.2)
• NA
• NVIDIA GPU Driver Version (valid for GPU only)
• Issue Type( questions)
Hi all
I am working on a project involving DeepStream’s deepstream_imagedata-multistream_cupy application, which builds on top of the deepstream-imagedata-multistream sample. This application allows for GPU-based image buffer access using CuPy arrays and supports multistream sources with uridecodebin.
I have introduced some GPU-heavy computations to simulate post-processing in the same thread as the DeepStream pipeline. Specifically, I added the following lines at line 125 to mimic the behavior:
A = cp.zeros((1000, 1000))
A + A
When this computation is executed, both video smoothness and AI inference performance degrade significantly.
Observations:
- Running the same computationally expensive task in a separate process (launched from a new command line) does not affect the video performance or AI inference.
- Attempting to perform the GPU computation:
- Inside a probe function: Causes the same issue.
- In a separate thread: Also causes degradation, similar to running it in the main pipeline.
This suggests that the issue lies in how the GPU tasks share resources, and I suspect it might be due to a lack of proper context switching or resource isolation.
Questions:
- Are there any CUDA primitives or configurations that can help separate GPU computations (e.g., post-processing) from the DeepStream pipeline tasks to avoid such interference?
- Is it possible to achieve this without offloading the computations to a separate process? If so, how?
- What are the recommended practices for handling GPU-intensive post-processing in DeepStream while maintaining video smoothness and inference quality?
I would greatly appreciate any insights or suggestions to resolve this issue.
Please find the code that I touched from nvidia examples.
#!/usr/bin/env python3
################################################################################
# SPDX-FileCopyrightText: Copyright (c) 2022-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
################################################################################
import sys
sys.path.append('../')
import gi
gi.require_version('Gst', '1.0')
from gi.repository import GLib, Gst
from ctypes import *
import sys
import math
from common.platform_info import PlatformInfo
from common.bus_call import bus_call
from common.FPS import PERF_DATA
import pyds
import argparse
import ctypes
import cupy as cp
import time
perf_data = None
MAX_DISPLAY_LEN = 64
PGIE_CLASS_ID_VEHICLE = 0
PGIE_CLASS_ID_BICYCLE = 1
PGIE_CLASS_ID_PERSON = 2
PGIE_CLASS_ID_ROADSIGN = 3
MUXER_OUTPUT_WIDTH = 1920
MUXER_OUTPUT_HEIGHT = 1080
MUXER_BATCH_TIMEOUT_USEC = 33000
TILED_OUTPUT_WIDTH = 1920
TILED_OUTPUT_HEIGHT = 1080
GST_CAPS_FEATURES_NVMM = "memory:NVMM"
pgie_classes_str= ["Vehicle", "TwoWheeler", "Person","RoadSign"]
# tiler_sink_pad_buffer_probe will extract metadata received on tiler src pad
# and modify the frame buffer using cupy
def tiler_sink_pad_buffer_probe(pad, info, u_data):
frame_number = 0
gst_buffer = info.get_buffer()
if not gst_buffer:
print("Unable to get GstBuffer ")
return
# Retrieve batch metadata from the gst_buffer
# Note that pyds.gst_buffer_get_nvds_batch_meta() expects the
# C address of gst_buffer as input, which is obtained with hash(gst_buffer)
batch_meta = pyds.gst_buffer_get_nvds_batch_meta(hash(gst_buffer))
l_frame = batch_meta.frame_meta_list
while l_frame is not None:
try:
# Note that l_frame.data needs a cast to pyds.NvDsFrameMeta
# The casting is done by pyds.NvDsFrameMeta.cast()
# The casting also keeps ownership of the underlying memory
# in the C code, so the Python garbage collector will leave
# it alone.
frame_meta = pyds.NvDsFrameMeta.cast(l_frame.data)
except StopIteration:
break
frame_number = frame_meta.frame_num
l_obj=frame_meta.obj_meta_list
num_rects = frame_meta.num_obj_meta
obj_counter = {
PGIE_CLASS_ID_VEHICLE:0,
PGIE_CLASS_ID_PERSON:0,
PGIE_CLASS_ID_BICYCLE:0,
PGIE_CLASS_ID_ROADSIGN:0
}
while l_obj is not None:
try:
# Casting l_obj.data to pyds.NvDsObjectMeta
obj_meta=pyds.NvDsObjectMeta.cast(l_obj.data)
except StopIteration:
break
obj_counter[obj_meta.class_id] += 1
try:
l_obj=l_obj.next
except StopIteration:
break
# Create dummy owner object to keep memory for the image array alive
owner = None
# Getting Image data using nvbufsurface
# the input should be address of buffer and batch_id
# Retrieve dtype, shape of the array, strides, pointer to the GPU buffer, and size of the allocated memory
data_type, shape, strides, dataptr, size = pyds.get_nvds_buf_surface_gpu(hash(gst_buffer), frame_meta.batch_id)
# dataptr is of type PyCapsule -> Use ctypes to retrieve the pointer as an int to pass into cupy
ctypes.pythonapi.PyCapsule_GetPointer.restype = ctypes.c_void_p
ctypes.pythonapi.PyCapsule_GetPointer.argtypes = [ctypes.py_object, ctypes.c_char_p]
# Get pointer to buffer and create UnownedMemory object from the gpu buffer
c_data_ptr = ctypes.pythonapi.PyCapsule_GetPointer(dataptr, None)
unownedmem = cp.cuda.UnownedMemory(c_data_ptr, size, owner)
# Create MemoryPointer object from unownedmem, at index 0
memptr = cp.cuda.MemoryPointer(unownedmem, 0)
# Create cupy array to access the image data. This array is in GPU buffer
n_frame_gpu = cp.ndarray(shape=shape, dtype=data_type, memptr=memptr, strides=strides, order='C')
# Initialize cuda.stream object for stream synchronization
stream = cp.cuda.stream.Stream(null=True) # Use null stream to prevent other cuda applications from making illegal memory access of buffer
# Modify the red channel to add blue tint to image
with stream:
# n_frame_gpu[:, :, 0] = 0.5 * n_frame_gpu[:, :, 0] + 0.5
#==============================new addition======================================
t0 = time.time()
while time.time()-t0 < 0.02: # do some heavy computation for 20 ms
for i in range(500):
A=cp.zeros((1000,1000))
A+A
# time.sleep(0.001) # comment above two lines and uncomment this to see improved performance
#=================================end============================================
stream.synchronize()
print("Frame Number=", frame_number, "Number of Objects=",num_rects,"Vehicle_count=",obj_counter[PGIE_CLASS_ID_VEHICLE],"Person_count=",obj_counter[PGIE_CLASS_ID_PERSON])
# Get frame rate through this probe
stream_index = "stream{0}".format(frame_meta.pad_index)
global perf_data
perf_data.update_fps(stream_index)
try:
l_frame = l_frame.next
except StopIteration:
break
return Gst.PadProbeReturn.OK
def cb_newpad(decodebin, decoder_src_pad, data):
print("In cb_newpad\n")
caps = decoder_src_pad.get_current_caps()
gststruct = caps.get_structure(0)
gstname = gststruct.get_name()
source_bin = data
features = caps.get_features(0)
# Need to check if the pad created by the decodebin is for video and not
# audio.
if (gstname.find("video") != -1):
# Link the decodebin pad only if decodebin has picked nvidia
# decoder plugin nvdec_*. We do this by checking if the pad caps contain
# NVMM memory features.
if features.contains("memory:NVMM"):
# Get the source bin ghost pad
bin_ghost_pad = source_bin.get_static_pad("src")
if not bin_ghost_pad.set_target(decoder_src_pad):
sys.stderr.write("Failed to link decoder src pad to source bin ghost pad\n")
else:
sys.stderr.write(" Error: Decodebin did not pick nvidia decoder plugin.\n")
def decodebin_child_added(child_proxy, Object, name, user_data):
print("Decodebin child added:", name, "\n")
if name.find("decodebin") != -1:
Object.connect("child-added", decodebin_child_added, user_data)
if "source" in name:
source_element = child_proxy.get_by_name("source")
if source_element.find_property('drop-on-latency') != None:
Object.set_property("drop-on-latency", True)
def create_source_bin(index, uri):
print("Creating source bin")
# Create a source GstBin to abstract this bin's content from the rest of the
# pipeline
bin_name = "source-bin-%02d" % index
print(bin_name)
nbin = Gst.Bin.new(bin_name)
if not nbin:
sys.stderr.write(" Unable to create source bin \n")
# Source element for reading from the uri.
# We will use decodebin and let it figure out the container format of the
# stream and the codec and plug the appropriate demux and decode plugins.
uri_decode_bin = Gst.ElementFactory.make("uridecodebin", "uri-decode-bin")
if not uri_decode_bin:
sys.stderr.write(" Unable to create uri decode bin \n")
# We set the input uri to the source element
uri_decode_bin.set_property("uri", uri)
# Connect to the "pad-added" signal of the decodebin which generates a
# callback once a new pad for raw data has beed created by the decodebin
uri_decode_bin.connect("pad-added", cb_newpad, nbin)
uri_decode_bin.connect("child-added", decodebin_child_added, nbin)
# We need to create a ghost pad for the source bin which will act as a proxy
# for the video decoder src pad. The ghost pad will not have a target right
# now. Once the decode bin creates the video decoder and generates the
# cb_newpad callback, we will set the ghost pad target to the video decoder
# src pad.
Gst.Bin.add(nbin, uri_decode_bin)
bin_pad = nbin.add_pad(Gst.GhostPad.new_no_target("src", Gst.PadDirection.SRC))
if not bin_pad:
sys.stderr.write(" Failed to add ghost pad in source bin \n")
return None
return nbin
def main(args):
global perf_data
perf_data = PERF_DATA(len(args))
number_sources = len(args)
# Standard GStreamer initialization
Gst.init(None)
# Create gstreamer elements */
# Create Pipeline element that will form a connection of other elements
print("Creating Pipeline \n ")
pipeline = Gst.Pipeline()
is_live = False
if not pipeline:
sys.stderr.write(" Unable to create Pipeline \n")
print("Creating streammux \n ")
# Create nvstreammux instance to form batches from one or more sources.
streammux = Gst.ElementFactory.make("nvstreammux", "Stream-muxer")
if not streammux:
sys.stderr.write(" Unable to create NvStreamMux \n")
pipeline.add(streammux)
for i in range(number_sources):
print("Creating source_bin ", i, " \n ")
uri_name = args[i]
if uri_name.find("rtsp://") == 0:
is_live = True
source_bin = create_source_bin(i, uri_name)
if not source_bin:
sys.stderr.write("Unable to create source bin \n")
pipeline.add(source_bin)
padname = "sink_%u" % i
sinkpad = streammux.request_pad_simple(padname)
if not sinkpad:
sys.stderr.write("Unable to create sink pad bin \n")
srcpad = source_bin.get_static_pad("src")
if not srcpad:
sys.stderr.write("Unable to create src pad bin \n")
srcpad.link(sinkpad)
print("Creating Pgie \n ")
pgie = Gst.ElementFactory.make("nvinfer", "primary-inference")
if not pgie:
sys.stderr.write(" Unable to create pgie \n")
# Add nvvidconv1 and filter1 to convert the frames to RGBA
# which is easier to work with in Python.
print("Creating nvvidconv1 \n ")
nvvidconv1 = Gst.ElementFactory.make("nvvideoconvert", "convertor1")
if not nvvidconv1:
sys.stderr.write(" Unable to create nvvidconv1 \n")
print("Creating filter1 \n ")
caps1 = Gst.Caps.from_string("video/x-raw(memory:NVMM), format=RGBA")
filter1 = Gst.ElementFactory.make("capsfilter", "filter1")
if not filter1:
sys.stderr.write(" Unable to get the caps filter1 \n")
filter1.set_property("caps", caps1)
print("Creating tiler \n ")
tiler = Gst.ElementFactory.make("nvmultistreamtiler", "nvtiler")
if not tiler:
sys.stderr.write(" Unable to create tiler \n")
print("Creating nvvidconv \n ")
nvvidconv = Gst.ElementFactory.make("nvvideoconvert", "convertor")
if not nvvidconv:
sys.stderr.write(" Unable to create nvvidconv \n")
print("Creating nvosd \n ")
nvosd = Gst.ElementFactory.make("nvdsosd", "onscreendisplay")
if not nvosd:
sys.stderr.write(" Unable to create nvosd \n")
if platform_info.is_platform_aarch64():
print("Creating nv3dsink \n")
sink = Gst.ElementFactory.make("nv3dsink", "nv3d-sink")
else:
print("Creating EGLSink \n")
sink = Gst.ElementFactory.make("nveglglessink", "nvvideo-renderer")
if not sink:
sys.stderr.write(" Unable to create egl sink \n")
if is_live:
print("Atleast one of the sources is live")
streammux.set_property('live-source', 1)
streammux.set_property('width', 1920)
streammux.set_property('height', 1080)
streammux.set_property('batch-size', number_sources)
streammux.set_property('batched-push-timeout', MUXER_BATCH_TIMEOUT_USEC)
pgie.set_property('config-file-path', "dstest_imagedata_cupy_config.txt")
pgie_batch_size = pgie.get_property("batch-size")
if (pgie_batch_size != number_sources):
print("WARNING: Overriding infer-config batch-size", pgie_batch_size, " with number of sources ",
number_sources, " \n")
pgie.set_property("batch-size", number_sources)
tiler_rows = int(math.sqrt(number_sources))
tiler_columns = int(math.ceil((1.0 * number_sources) / tiler_rows))
tiler.set_property("rows", tiler_rows)
tiler.set_property("columns", tiler_columns)
tiler.set_property("width", TILED_OUTPUT_WIDTH)
tiler.set_property("height", TILED_OUTPUT_HEIGHT)
sink.set_property("sync", 0)
sink.set_property("qos", 0)
print("Adding elements to Pipeline \n")
pipeline.add(pgie)
pipeline.add(tiler)
pipeline.add(nvvidconv)
pipeline.add(filter1)
pipeline.add(nvvidconv1)
pipeline.add(nvosd)
pipeline.add(sink)
print("Linking elements in the Pipeline \n")
streammux.link(pgie)
pgie.link(nvvidconv1)
nvvidconv1.link(filter1)
filter1.link(tiler)
tiler.link(nvvidconv)
nvvidconv.link(nvosd)
nvosd.link(sink)
# create an event loop and feed gstreamer bus mesages to it
loop = GLib.MainLoop()
bus = pipeline.get_bus()
bus.add_signal_watch()
bus.connect("message", bus_call, loop)
tiler_sink_pad = tiler.get_static_pad("sink")
if not tiler_sink_pad:
sys.stderr.write(" Unable to get src pad \n")
else:
tiler_sink_pad.add_probe(Gst.PadProbeType.BUFFER, tiler_sink_pad_buffer_probe, 0)
# perf callback function to print fps every 5 sec
GLib.timeout_add(5000, perf_data.perf_print_callback)
# List the sources
print("Now playing...")
for i, source in enumerate(args):
print(i, ": ", source)
print("Starting pipeline \n")
# start play back and listed to events
pipeline.set_state(Gst.State.PLAYING)
try:
loop.run()
except:
pass
# cleanup
print("Exiting app\n")
pipeline.set_state(Gst.State.NULL)
def parse_args():
parser = argparse.ArgumentParser(prog="deepstream_imagedata-multistream_cupy.py",
description="deepstream-imagedata-multistream-cupy takes multiple URI streams as input" \
" and retrieves the image buffer from GPU as a cupy array for in-place modification")
parser.add_argument(
"-i",
"--input",
help="Path to input streams",
nargs="+",
metavar="URIs",
default=["a"],
required=True,
)
args = parser.parse_args()
stream_paths = args.input
return stream_paths
if __name__ == '__main__':
platform_info = PlatformInfo()
if platform_info.is_integrated_gpu():
sys.stderr.write ("\nThis app is not currently supported on integrated GPU. Exiting...\n\n\n\n")
sys.exit(1)
stream_paths = parse_args()
sys.exit(main(stream_paths))
Essentially we are looking for some solutions to seperate heavy cupy computaion from video streaming to have a smooth video and analytics experience.
Note: We cannot use triton or nvinfer to do our inference. We are doing our inference using cupy extraction.
