GPU视频编解码:X86 DeepStream 视频编解码入门(三)

一.设备硬件信息

1.X86 auto 云端主机

bash 复制代码
video) (base) root@autodl-container-2b344a8a9f-e5fa6b67:~/autodl-tmp/deepstream_libraries/encode_video# cat /etc/os-release 
PRETTY_NAME="Ubuntu 22.04.3 LTS"
NAME="Ubuntu"
VERSION_ID="22.04"
VERSION="22.04.3 LTS (Jammy Jellyfish)"
VERSION_CODENAME=jammy
ID=ubuntu
ID_LIKE=debian
HOME_URL="https://www.ubuntu.com/"
SUPPORT_URL="https://help.ubuntu.com/"
BUG_REPORT_URL="https://bugs.launchpad.net/ubuntu/"
PRIVACY_POLICY_URL="https://www.ubuntu.com/legal/terms-and-policies/privacy-policy"
UBUNTU_CODENAME=jammy

2.python3 版本 必须得是3.10

bash 复制代码
(video) (base) root@autodl-container-2b344a8a9f-e5fa6b67:~/autodl-tmp/deepstream_libraries/encode_video# python3
Python 3.10.0 (default, Mar  3 2022, 09:58:08) [GCC 7.5.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> eixt()

3.cuda 版本必须大于12

cpp 复制代码
(video) (base) root@autodl-container-2b344a8a9f-e5fa6b67:~/autodl-tmp/deepstream_libraries/encode_video# nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2023 NVIDIA Corporation
Built on Mon_Apr__3_17:16:06_PDT_2023
Cuda compilation tools, release 12.1, V12.1.105
Build cuda_12.1.r12.1/compiler.32688072_0

二. 生态环境安装

1.源码仓库clone:https://github.com/NVIDIA-AI-IOT/deepstream_libraries.git

2.whl文件下载:以下方式都行

cpp 复制代码
wget --content-disposition 'https://api.ngc.nvidia.com/v2/resources/org/nvidia/deepstream/7.1/files?redirect=true&path=deepstream_libraries-1.1-cp310-cp310-linux_x86_64.whl' -O deepstream_libraries-1.1-cp310-cp310-linux_x86_64.whl
cpp 复制代码
curl -LO 'https://api.ngc.nvidia.com/v2/resources/org/nvidia/deepstream/7.1/files?redirect=true&path=deepstream_libraries-1.1-cp310-cp310-linux_x86_64.whl'

3.安装 :必须是python3.10

cpp 复制代码
conda create -n video python=3.10 
conda activate video
cpp 复制代码
(video) root@autodl-container-2b344a8a9f-e5fa6b67:~/autodl-tmp/deepstream_libraries# pip3 install de
decode_video/                                          deepstream_libraries-1.1-cp310-cp310-linux_x86_64.whl
(video) root@autodl-container-2b344a8a9f-e5fa6b67:~/autodl-tmp/deepstream_libraries# pip3 install deepstream_libraries-1.1-cp310-cp310-linux_x86_64.whl 
Looking in indexes: http://mirrors.aliyun.com/pypi/simple
Processing ./deepstream_libraries-1.1-cp310-cp310-linux_x86_64.whl
Collecting cvcuda-cu12@ https://github.com/CVCUDA/CV-CUDA/releases/download/v0.10.1-beta/cvcuda_cu12-0.10.1b0-cp310-cp310-linux_x86_64.whl (from deepstream-libraries==1.1)
  Downloading https://github.com/CVCUDA/CV-CUDA/releases/download/v0.10.1-beta/cvcuda_cu12-0.10.1b0-cp310-cp310-linux_x86_64.whl (124.8 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 124.8/124.8 MB 9.1 MB/s eta 0:00:00
Collecting PyNvVideoCodec==1.0.1 (from deepstream-libraries==1.1)
  Downloading http://mirrors.aliyun.com/pypi/packages/24/ce/3789d47a2336319a02563e6071eb85eba643e9f9d22bdcf3415d0d2a0aa3/PyNvVideoCodec-1.0.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (56.8 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 56.8/56.8 MB 13.8 MB/s eta 0:00:00
Collecting nvidia-nvimgcodec-cu12==0.3.0.5 (from deepstream-libraries==1.1)
  Downloading http://mirrors.aliyun.com/pypi/packages/67/6d/749eb173622a0f269c967abc0f946970899e90970427b1b8446e8b7ba8d1/nvidia_nvimgcodec_cu12-0.3.0.5-py3-none-manylinux2014_x86_64.whl (10.9 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 10.9/10.9 MB 13.8 MB/s eta 0:00:00
Collecting numpy<2.0.0,>=1.23.5 (from cvcuda-cu12@ https://github.com/CVCUDA/CV-CUDA/releases/download/v0.10.1-beta/cvcuda_cu12-0.10.1b0-cp310-cp310-linux_x86_64.whl->deepstream-libraries==1.1)
  Downloading http://mirrors.aliyun.com/pypi/packages/4b/d7/ecf66c1cd12dc28b4040b15ab4d17b773b87fa9d29ca16125de01adb36cd/numpy-1.26.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (18.2 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 18.2/18.2 MB 15.3 MB/s eta 0:00:00
Installing collected packages: PyNvVideoCodec, nvidia-nvimgcodec-cu12, numpy, cvcuda-cu12, deepstream-libraries
Successfully installed PyNvVideoCodec-1.0.1 cvcuda-cu12-0.10.1b0 deepstream-libraries-1.1 numpy-1.26.4 nvidia-nvimgcodec-cu12-0.3.0.5
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager, possibly rendering your system unusable. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv. Use the --root-user-action option if you know what you are doing and want to suppress this warning.
(video) root@autodl-container-2b344a8a9f-e5fa6b67:~/autodl-tmp/deepstream_libraries# 

三.依赖包讲解

1.CVCUDA是字节和NVIDIA开发的基于CUDA进行视频图像处理工具,代替OpenCV

2.PyNvVideoCodec继承自上一节用到的VPF框架:

cpp 复制代码
pip install PyNvVideoCodec==1.0.1
  1. nvidia-nvimgcodec这是处理图片的工具

4.pycuda提供python cuda 接口

四.视频编码实例

1.视频编码:成功将mp4进行h264编码

cpp 复制代码
(video) (base) root@autodl-container-2b344a8a9f-e5fa6b67:~/autodl-tmp/deepstream_libraries# cd encode_video/
(video) (base) root@autodl-container-2b344a8a9f-e5fa6b67:~/autodl-tmp/deepstream_libraries/encode_video# python3 encode.py --gpu_id 0 --raw_file_path ../assets/videos/pexels-chiel-slotman-4423925-1920x1080-25fps.mp4  --encoded_file_path output.h264 --size 1920x1080 --format nv12 --codec h264 --config_file ../assets/configs/encode_config.json
context fetched=0x562094010510
stream created=0x562094977a90
(video) (base) root@autodl-container-2b344a8a9f-e5fa6b67:~/autodl-tmp/deepstream_libraries/encode_video# ls
README.md  encode.py  output.h264
(video) (base) root@autodl-container-2b344a8a9f-e5fa6b67:~/autodl-tmp/deepstream_libraries/encode_video# 

2.编码前后文件大小对比,降低了75%+存储空间

cpp 复制代码
(video) (base) root@autodl-container-2b344a8a9f-e5fa6b67:~/autodl-tmp/deepstream_libraries/encode_video# du -sh output.h264 
1.1M    output.h264
(video) (base) root@autodl-container-2b344a8a9f-e5fa6b67:~/autodl-tmp/deepstream_libraries/encode_video# du -sh ../assets/videos/pexels-chiel-slotman-4423925-1920x1080-25fps.mp4
4.6M    ../assets/videos/pexels-chiel-slotman-4423925-1920x1080-25fps.mp4
(video) (base) root@autodl-container-2b344a8a9f-e5fa6b67:~/autodl-tmp/deepstream_libraries/encode_video# 

3.代码讲解(非常简洁,比前面几个系列硬编解码都nice!)

导入依赖

python 复制代码
import io
import sys
import json
import argparse
sys.path.append('../')
from pathlib import Path
import PyNvVideoCodec as nvc
from common.nvc_utils import FetchGPUFrame, FetchCPUFrame, AppFrame

total_num_frames = 1000

函数参数列表解析

python 复制代码
def encode(gpu_id, dec_file_path, enc_file_path, width, height, fmt, config_params):
    """
                This function illustrates encoding of frames using CUDA device buffers as input.

                The application reads the image data from file and loads it to CUDA input buffers using
                FetchGPUFrame(). The encoder subsequently copies the CUDA buffers and submits them to NVENC hardware
                for encoding as part of Encode() function. Video memory buffer allocated
                by the application to get the NVENC hardware output. This application copies the NVENC output
                from video memory buffer to host memory buffer in order to dump to a file, but this
                is not needed if application choose to use it in some other way.

                Parameters:
                    - gpu_id (int): Ordinal of GPU to use [Parameter not in use]
                    - dec_file_path (str): Path to
                file to be decoded
                    - enc_file_path (str): Path to output file into which raw frames are stored
                    - width (int): width of encoded frame
                    - height (int): height of encoded frame
                    - fmt (str) : surface format string in uppercase, for e.g. NV12
                    - config_params(key value pairs) : key value pairs providing fine-grained control on encoding

                Returns: - None.

函数调用实例

python 复制代码
Example:
                >>> encode(0, "path/to/input/yuv/file","path/to/output/elementary/bitstream",1920,1080,"NV12")
                Encode 1080p NV12 raw YUV into elementary bitstream using H.264 codec and P4 preset

pipline说明

python 复制代码
 with open(dec_file_path, "rb") as decFile, open(enc_file_path, "wb") as encFile:
        nvenc = nvc.CreateEncoder(width, height, fmt, False, **config_params)  # create encoder object
        input_frame_list = list([AppFrame(width, height, fmt) for x in range(1, 5)])
        for input_gpu_frame in FetchGPUFrame(input_frame_list,
                                             FetchCPUFrame(decFile, input_frame_list[0].frameSize),
                                             total_num_frames):
            bitstream = nvenc.Encode(input_gpu_frame)  # encode frame one by one
            bitstream = bytearray(bitstream)
            encFile.write(bitstream)

        bitstream = nvenc.EndEncode()  # flush encoder queue
        bitstream = bytearray(bitstream)
        encFile.write(bitstream)

五.视频解码

1.解码cli命令

cpp 复制代码
video) (base) root@autodl-container-2b344a8a9f-e5fa6b67:~/autodl-tmp/deepstream_libraries/decode_video# python3 decode.py --gpu_id 0 --encoded_file_path ../assets/videos/pexels-chiel-slotman-4423925-1920x1080-25fps.mp4 --raw_file_path output.yuv --use_device_memory 1
[INFO ][00:13:25] Media format: QuickTime / MOV (mov,mp4,m4a,3gp,3g2,mj2)
[NULL @ 0x562dad250d40] No codec provided to avcodec_open2()
[WARN ][00:13:25] ChromaFormat not recognized. Assuming 420
FPS =  25.0
Session Initialization Time: 21 ms 
Session Deinitialization Time: 226 ms 

2.依赖包

cpp 复制代码
import sys
import argparse
import numpy as np

sys.path.append('../')
from pathlib import Path
import pycuda.driver as cuda
import PyNvVideoCodec as nvc
from common.nvc_utils import cast_address_to_1d_bytearray

3.创建解码器和缓冲空间

cpp 复制代码
    nv_dmx = nvc.CreateDemuxer(filename=enc_file_path)
    nv_dec = nvc.CreateDecoder(gpuid=0,
                               codec=nv_dmx.GetNvCodecId(),
                               cudacontext=0,
                               cudastream=0,
                               usedevicememory=use_device_memory)

4.pipline

cpp 复制代码
# 以写入模式打开输出文件
    with open(dec_file_path, "wb") as decFile:
        # 遍历解复用器,逐个获取数据包(packet)
        for packet in nv_dmx:
            # 送入解码器进行解码,解码器返回多个解码后的帧
            for decoded_frame in nv_dec.Decode(packet):
                # 对于 NV12 格式,解码器会返回两个平面(Y 平面和 UV 平面)
                if not seq_triggered:
                    decoded_frame_size = nv_dec.GetFrameSize()  # 获取解码后帧的字节大小
                    raw_frame = np.ndarray(shape=decoded_frame_size, dtype=np.uint8)  # 申请内存存储解码帧
                    seq_triggered = True
                
                # 获取 Y 平面的设备地址
                luma_base_addr = decoded_frame.GetPtrToPlane(0)
                
                if use_device_memory:
                    # 如果使用 GPU 设备内存,需要手动从 GPU 复制到 CPU
                    cuda.memcpy_dtoh(raw_frame, luma_base_addr)  # 设备到主机拷贝
                    bits = bytearray(raw_frame)  # 转换为字节流
                    decFile.write(bits)  # 写入文件
                else:
                    # 如果使用的是主机内存,则直接转换地址为 1D 数组并写入文件
                    new_array = cast_address_to_1d_bytearray(
                        base_address=luma_base_addr,
                        size=decoded_frame.framesize()
                    )
                    decFile.write(bytearray(new_array))
相关推荐
nice_lcj5202 小时前
MySQL 主从复制与读写分离:从库怎么跟着主库走
数据库·mysql·面试·binlog·读写分离·主从复制
tianyuanwo3 小时前
PostgreSQL 部署运维实战总结:从安装踩坑到生产可用
运维·数据库·postgresql
数据工匠老o3 小时前
数据库没死,业务卡住了:行锁等待、元数据锁、长事务的排查路径
数据库·算法·架构
Decay__3 小时前
SQLServer 优化案例集 —— 慢 SQL 优化、索引重构与参数调整收益对比
运维·数据库·sql·mysql·架构
CoderYanger4 小时前
Java EE 进阶:9.MyBatis 操作数据库(进阶)(第二弹)
java·数据库·面试·职场和发展·java-ee·tomcat·mybatis
龙腾AI白云4 小时前
具身智能的空间导航与未知环境探索
大数据·数据库·人工智能·机器学习·知识图谱
ClouGence5 小时前
DBeaver太重,Navicat付费?试试这4款好用的开源数据库管理工具!
数据库·开源
字节跳动数据库5 小时前
火山引擎缓存数据库:如何构建超大规模缓存集群的无感弹性能力
数据库·人工智能·后端
SelectDB5 小时前
多模态检索用什么数据库?向量、全文与 SQL 混合检索选型指南
数据库