Continuous growth of CPU and memory

[liujiachang]

When I used SampleDecodeMultiThread. py for testing, the CPU usage and memory usage of the program continued to increase. Initially, it only occupied 300% CPU and 1.8G mem. After running for about ten minutes, it occupied 1000% CPU and 2.3G mem. I slightly modified the test script and read fifty video streams at the same time. I don't think it should consume so much resources.
Here is my code：

import PyNvCodec as nvc
# import numpy as np
import argparse
from threading import Thread
import threading
# import cv2
import time


class Worker(Thread):
    def __init__(self, gpuID, encFile):
        Thread.__init__(self)
        self.st = time.time()
        self.pt = time.time()
        # Create Decoder with given CUDA context & stream.
        self.nvDec = nvc.PyNvDecoder(encFile, gpuID)

        self.width, self.height = self.nvDec.Width(), self.nvDec.Height()

        # Determine colorspace conversion parameters.
        # Some video streams don't specify these parameters so default values
        # are most widespread bt601 and mpeg.
        cspace, crange = self.nvDec.ColorSpace(), self.nvDec.ColorRange()
        if nvc.ColorSpace.UNSPEC == cspace:
            cspace = nvc.ColorSpace.BT_601
        if nvc.ColorRange.UDEF == crange:
            crange = nvc.ColorRange.MPEG
        self.cc_ctx = nvc.ColorspaceConversionContext(cspace, crange)
        # print("Color space: ", str(cspace))
        # print("Color range: ", str(crange))

        # Initialize colorspace conversion chain
        if self.nvDec.ColorSpace() != nvc.ColorSpace.BT_709:
            self.nvYuv = nvc.PySurfaceConverter(
                self.width,
                self.height,
                self.nvDec.Format(),
                nvc.PixelFormat.YUV420,
                0
            )
        else:
            self.nvYuv = None

        if self.nvYuv:
            self.nvCvt = nvc.PySurfaceConverter(
                self.width,
                self.height,
                self.nvYuv.Format(),
                nvc.PixelFormat.BGR,
                0
            )
        else:
            self.nvCvt = nvc.PySurfaceConverter(
                self.width,
                self.height,
                self.nvDec.Format(),
                nvc.PixelFormat.BGR,
                0
            )
        # # resize
        # self.nvRes = nvc.PySurfaceResizer(
        #     hwidth, hheight, self.nvCvt.Format(), 0
        # )
        # self.nvDwn = nvc.PySurfaceDownloader(
        #     hwidth, hheight, self.nvRes.Format(), 0
        # )
        self.nvDwn = nvc.PySurfaceDownloader(
            self.width, self.height, self.nvCvt.Format(), 0
        )
        self.num_frame = 0

    def run(self):
        try:
            while True:
                try:
                    self.rawSurface = self.nvDec.DecodeSingleSurface()
                    if self.rawSurface.Empty():
                        print("No more video frames")
                        break
                except nvc.HwResetException:
                    print("Continue after HW decoder was reset")
                    continue
                if 0 == self.num_frame % 5:
                    if self.nvYuv:
                        self.yuvSurface = self.nvYuv.Execute(self.rawSurface, self.cc_ctx)
                        self.cvtSurface = self.nvCvt.Execute(self.yuvSurface, self.cc_ctx)
                    else:
                        self.cvtSurface = self.nvCvt.Execute(self.rawSurface, self.cc_ctx)
                    if self.cvtSurface.Empty():
                        print("Failed to do color conversion")
                        break

                    # self.resSurface = self.nvRes.Execute(self.cvtSurface)
                    # if self.resSurface.Empty():
                    #     print("Failed to resize surface")
                    #     break

                    # self.rawFrame = np.ndarray(
                    #     shape=(self.cvtSurface.HostSize()), dtype=np.uint8
                    # )
                    # success = self.nvDwn.DownloadSingleSurface(
                    #     self.cvtSurface, self.rawFrame
                    # )
                    #
                    # if not (success):
                    #     print("Failed to download surface")
                    #     break
                    # print(self.rawFrame.shape)

                    # self.rawFrame = self.rawFrame.reshape((self.height, self.width, 3))
                    # cv2.imwrite('test.png', self.rawFrame)
                    # break

                    # print(f"Thread {self.ident} at frame {self.num_frame}")
                if time.time() - self.pt > 5:
                    self.pt = time.time()
                    print(f"Thread {threading.current_thread().name} fps {self.num_frame / (time.time() - self.st)}")
                self.num_frame += 1
        except Exception as e:
            print(getattr(e, "message", str(e)))


def create_threads(gpu_id, input_file, num_threads):

    thread_pool = []
    for i in range(0, num_threads):
        thread = Worker(gpu_id, input_file)
        thread.start()
        thread_pool.append(thread)

    for thread in thread_pool:
        thread.join()


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description='frame decode')
    parser.add_argument('--nums', type=int, default=50, help="stream nums")
    parser.add_argument('--gpu', type=int, default=0, help="gpu id")
    args = parser.parse_args()
    stream = '*******'  #  This is my video streaming address, sorry I cannot disclose it.

    create_threads(args.gpu, stream, args.nums)

[david-molnar-oculai]

We're using the decoder and encoder and run them continuously to process a stream of 4 second clips. The memory and cpu doesn't increase, it stays constant. We used https://github.com/NVIDIA/VideoProcessingFramework/blob/master/samples/SamplePyTorch.py as a base. Also, more streams are processed on the same machine, using different threads in the same process.