一、硬件编码核心流程
-
硬件设备初始化
// 创建CUDA硬件设备上下文 AVBufferRef *hw_device_ctx = NULL; av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_CUDA, NULL, NULL, 0);// 绑定硬件设备到编码器上下文 codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
-
编码器选择与参数配置
// 查找NVIDIA H.264硬件编码器 const AVCodec *encoder = avcodec_find_encoder_by_name("h264_nvenc");// 设置编码参数(分辨率、码率、GOP等) codec_ctx->width = 1920; codec_ctx->height = 1080; codec_ctx->bit_rate = 5000000; codec_ctx->time_base = (AVRational){1, 30}; codec_ctx->pix_fmt = AV_PIX_FMT_CUDA; // 指定硬件像素格式// 设置编码预设参数(NVIDIA专用) av_opt_set(codec_ctx->priv_data, "preset", "llhp", 0); // 低延迟高性能
-
硬件帧内存分配
// 创建硬件帧并绑定GPU内存 AVFrame *hw_frame = av_frame_alloc(); hw_frame->format = codec_ctx->pix_fmt; hw_frame->width = codec_ctx->width; hw_frame->height = codec_ctx->height; av_hwframe_get_buffer(codec_ctx->hw_frames_ctx, hw_frame, 0);
-
编码数据流处理
// 送入硬件编码器 avcodec_send_frame(codec_ctx, hw_frame);// 接收编码后的数据包 AVPacket *pkt = av_packet_alloc(); while (avcodec_receive_packet(codec_ctx, pkt) >= 0) {av_interleaved_write_frame(output_ctx, pkt); // 写入输出文件av_packet_unref(pkt); }
二、跨平台适配示例
硬件平台 | 编码器名称 | 像素格式 | 初始化函数 |
---|---|---|---|
NVIDIA GPU | h264_nvenc | AV_PIX_FMT_CUDA | av_hwdevice_ctx_create(..., AV_HWDEVICE_TYPE_CUDA) |
Intel QSV | h264_qsv | AV_PIX_FMT_QSV | av_hwdevice_ctx_create(..., AV_HWDEVICE_TYPE_QSV) |
AMD AMF | h264_amf | AV_PIX_FMT_D3D11 | av_hwdevice_ctx_create(..., AV_HWDEVICE_TYPE_D3D11VA) |
三、关键问题解决
-
编码器初始化失败
- 检查FFmpeg编译时是否启用对应硬件加速选项(如
--enable-nvenc
、--enable-libmfx
) - 确认硬件驱动版本与FFmpeg兼容性
- 检查FFmpeg编译时是否启用对应硬件加速选项(如
-
CPU-GPU内存拷贝开销优化
// 使用hwupload滤镜直接上传数据至GPU AVFilterContext *upload_filter; const AVFilter *hwupload = avfilter_get_by_name("hwupload"); avfilter_graph_create_filter(&upload_filter, hwupload, "upload", NULL, NULL, filter_graph);
四、资源释放
av_buffer_unref(&hw_device_ctx); // 释放硬件设备上下文
avcodec_free_context(&codec_ctx); // 释放编码器上下文
av_frame_free(&hw_frame); // 释放硬件帧
av_packet_free(&pkt); // 释放数据包
五、编译依赖
- NVIDIA平台:需安装CUDA Toolkit,编译时添加
--enable-cuda --enable-nvenc
- Intel平台:需安装Intel Media SDK,编译时添加
--enable-libmfx
六、硬件编码示例代码
支持从本地YUV文件读取数据、GPU加速编码并输出H.264视频流到MP4文件。
#include <libavcodec/avcodec.h>
#include <libavformat/avformat.h>
#include <libavutil/imgutils.h>
#include <libavutil/hwcontext.h>
#include <libswscale/swscale.h>#define INPUT_FILE "input.yuv"
#define OUTPUT_FILE "output.mp4"
#define WIDTH 1280
#define HEIGHT 720
#define FRAME_RATE 30int main() {AVFormatContext *fmt_ctx = NULL;AVCodecContext *enc_ctx = NULL;AVBufferRef *hw_device_ctx = NULL;SwsContext *sws_ctx = NULL;int ret = 0;// 1. 初始化硬件设备上下文ret = av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_CUDA, NULL, NULL, 0); // :ml-citation{ref="1,3" data="citationList"}if (ret < 0) {fprintf(stderr, "Failed to create CUDA device\n");goto cleanup;}// 2. 打开输出文件并配置封装格式avformat_alloc_output_context2(&fmt_ctx, NULL, NULL, OUTPUT_FILE); // :ml-citation{ref="8" data="citationList"}if (!fmt_ctx) {fprintf(stderr, "Failed to create output context\n");ret = -1;goto cleanup;}// 3. 查找并配置硬件编码器const AVCodec *encoder = avcodec_find_encoder_by_name("h264_nvenc"); // :ml-citation{ref="1,4" data="citationList"}if (!encoder) {fprintf(stderr, "NVENC encoder not found\n");ret = -1;goto cleanup;}enc_ctx = avcodec_alloc_context3(encoder);enc_ctx->width = WIDTH;enc_ctx->height = HEIGHT;enc_ctx->time_base = (AVRational){1, FRAME_RATE};enc_ctx->pix_fmt = AV_PIX_FMT_CUDA; // 硬件像素格式 :ml-citation{ref="3,5" data="citationList"}enc_ctx->bit_rate = 4000000; // 4Mbps码率 :ml-citation{ref="5,8" data="citationList"}enc_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx); // 绑定设备 :ml-citation{ref="3,6" data="citationList"}// 4. 打开编码器并添加视频流if ((ret = avcodec_open2(enc_ctx, encoder, NULL)) < 0) { // :ml-citation{ref="4" data="citationList"}fprintf(stderr, "Failed to open encoder\n");goto cleanup;}AVStream *stream = avformat_new_stream(fmt_ctx, NULL);avcodec_parameters_from_context(stream->codecpar, enc_ctx);// 5. 打开输出文件并写入头信息if (!(fmt_ctx->oformat->flags & AVFMT_NOFILE)) {ret = avio_open(&fmt_ctx->pb, OUTPUT_FILE, AVIO_FLAG_WRITE); // :ml-citation{ref="8" data="citationList"}if (ret < 0) {fprintf(stderr, "Failed to open output file\n");goto cleanup;}}avformat_write_header(fmt_ctx, NULL); // :ml-citation{ref="8" data="citationList"}// 6. 初始化YUV到CUDA格式转换器sws_ctx = sws_getContext(WIDTH, HEIGHT, AV_PIX_FMT_YUV420P,WIDTH, HEIGHT, AV_PIX_FMT_CUDA,SWS_BILINEAR, NULL, NULL, NULL); // :ml-citation{ref="4,7" data="citationList"}// 7. 准备输入帧和硬件帧AVFrame *yuv_frame = av_frame_alloc();yuv_frame->width = WIDTH;yuv_frame->height = HEIGHT;yuv_frame->format = AV_PIX_FMT_YUV420P;av_frame_get_buffer(yuv_frame, 0);AVFrame *hw_frame = av_frame_alloc();hw_frame->format = enc_ctx->pix_fmt;hw_frame->width = WIDTH;hw_frame->height = HEIGHT;av_hwframe_get_buffer(enc_ctx->hw_frames_ctx, hw_frame, 0); // :ml-citation{ref="3,6" data="citationList"}// 8. 处理每一帧数据FILE *yuv_file = fopen(INPUT_FILE, "rb");AVPacket *pkt = av_packet_alloc();for (int i = 0; i < 100; i++) { // 编码100帧测试// 从YUV文件读取数据fread(yuv_frame->data, 1, WIDTH*HEIGHT, yuv_file); // Y分量fread(yuv_frame->data:ml-citation{ref="1" data="citationList"}, 1, (WIDTH/2)*(HEIGHT/2), yuv_file); // U分量fread(yuv_frame->data:ml-citation{ref="2" data="citationList"}, 1, (WIDTH/2)*(HEIGHT/2), yuv_file); // V分量// 转换到硬件帧sws_scale(sws_ctx, (const uint8_t**)yuv_frame->data, yuv_frame->linesize,0, HEIGHT, hw_frame->data, hw_frame->linesize); // :ml-citation{ref="4,7" data="citationList"}// 编码并写入文件avcodec_send_frame(enc_ctx, hw_frame);while (avcodec_receive_packet(enc_ctx, pkt) >= 0) {av_packet_rescale_ts(pkt, enc_ctx->time_base, stream->time_base);av_interleaved_write_frame(fmt_ctx, pkt); // :ml-citation{ref="8" data="citationList"}av_packet_unref(pkt);}}cleanup:// 9. 释放所有资源if (yuv_file) fclose(yuv_file);if (sws_ctx) sws_freeContext(sws_ctx);av_packet_free(&pkt);av_frame_free(&yuv_frame);av_frame_free(&hw_frame);if (fmt_ctx) avformat_free_context(fmt_ctx);avcodec_free_context(&enc_ctx);av_buffer_unref(&hw_device_ctx);return ret;
}
关键代码说明:
组件 | 功能说明 | 依赖项 |
---|---|---|
hw_device_ctx | 管理CUDA设备上下文,用于GPU内存分配和硬件加速操作 | CUDA驱动和NVENC支持 |
sws_ctx | 将CPU端的YUV420P数据转换为GPU端的CUDA格式(如NV12) | libswscale库 |
av_hwframe_get_buffer | 直接从GPU显存分配帧内存,避免CPU-GPU内存拷贝 | FFmpeg硬件帧支持 |
avcodec_send_frame | 将硬件帧送入编码器队列,触发异步编码操作 | 编码器线程模型 |