Recherche avancée

Sur d’autres sites (9842)

• FFMPEG using AV_PIX_FMT_D3D11 gives "Error registering the input resource" from NVENC

  13 novembre 2024, par nbabcock

  Input frames start on the GPU as ID3D11Texture2D pointers.

  


  I encode them to H264 using FFMPEG + NVENC. NVENC works perfectly if I download the textures to CPU memory as format AV_PIX_FMT_BGR0, but I'd like to cut out the CPU texture download entirely, and pass the GPU memory pointer directly into the encoder in native format. I write frames like this :

  


  int write_gpu_video_frame(ID3D11Texture2D* gpuTex, AVFormatContext* oc, OutputStream* ost) {
    AVFrame *hw_frame = ost->hw_frame;

    printf("gpuTex address = 0x%x\n", &gpuTex);

    hw_frame->data[0] = (uint8_t *) gpuTex;
    hw_frame->data[1] = (uint8_t *) (intptr_t) 0;
    hw_frame->pts     = ost->next_pts++;

    return write_frame(oc, ost->enc, ost->st, hw_frame);
    // write_frame is identical to sample code in ffmpeg repo
}


  


  Running the code with this modification gives the following error :

  


  gpuTex address = 0x4582f6d0
[h264_nvenc @ 00000191233e1bc0] Error registering an input resource: invalid call (9):
[h264_nvenc @ 00000191233e1bc0] Could not register an input HW frame
Error sending a frame to the encoder: Unknown error occurred


  


  ---

  


  Here's some supplemental code used in setting up and configuring the hw context and encoder :

  


  /* A few config flags */
#define ENABLE_NVENC TRUE
#define USE_D3D11 TRUE // Skip downloading textures to CPU memory and send it straight to NVENC


  


  /* Init hardware frame context */
static int set_hwframe_ctx(AVCodecContext* ctx, AVBufferRef* hw_device_ctx) {
    AVBufferRef*       hw_frames_ref;
    AVHWFramesContext* frames_ctx = NULL;
    int                err        = 0;

    if (!(hw_frames_ref = av_hwframe_ctx_alloc(hw_device_ctx))) {
        fprintf(stderr, "Failed to create HW frame context.\n");
        throw;
    }
    frames_ctx                    = (AVHWFramesContext*) (hw_frames_ref->data);
    frames_ctx->format            = AV_PIX_FMT_D3D11;
    frames_ctx->sw_format         = AV_PIX_FMT_NV12;
    frames_ctx->width             = STREAM_WIDTH;
    frames_ctx->height            = STREAM_HEIGHT;
    //frames_ctx->initial_pool_size = 20;
    if ((err = av_hwframe_ctx_init(hw_frames_ref)) < 0) {
        fprintf(stderr, "Failed to initialize hw frame context. Error code: %s\n", av_err2str(err));
        av_buffer_unref(&hw_frames_ref);
        throw;
    }
    ctx->hw_frames_ctx = av_buffer_ref(hw_frames_ref);
    if (!ctx->hw_frames_ctx)
        err = AVERROR(ENOMEM);

    av_buffer_unref(&hw_frames_ref);
    return err;
}


  


  /* Add an output stream. */
static void add_video_stream(
    OutputStream* ost,
    AVFormatContext* oc,
    const AVCodec** codec,
    enum AVCodecID  codec_id,
    int width,
    int height
) {
    AVCodecContext* c;
    int             i;
    bool            nvenc = false;

    /* find the encoder */
    if (ENABLE_NVENC) {
        printf("Getting nvenc encoder\n");
        *codec = avcodec_find_encoder_by_name("h264_nvenc");
        nvenc  = true;
    }
    
    if (!ENABLE_NVENC || *codec == NULL) {
        printf("Getting standard encoder\n");
        avcodec_find_encoder(codec_id);
        nvenc = false;
    }
    if (!(*codec)) {
        fprintf(stderr, "Could not find encoder for '%s'\n",
                avcodec_get_name(codec_id));
        exit(1);
    }

    ost->st = avformat_new_stream(oc, NULL);
    if (!ost->st) {
        fprintf(stderr, "Could not allocate stream\n");
        exit(1);
    }
    ost->st->id = oc->nb_streams - 1;
    c           = avcodec_alloc_context3(*codec);
    if (!c) {
        fprintf(stderr, "Could not alloc an encoding context\n");
        exit(1);
    }
    ost->enc = c;

    printf("Using video codec %s\n", avcodec_get_name(codec_id));

    c->codec_id = codec_id;
    c->bit_rate = 4000000;
    /* Resolution must be a multiple of two. */
    c->width  = STREAM_WIDTH;
    c->height = STREAM_HEIGHT;
    /* timebase: This is the fundamental unit of time (in seconds) in terms
        * of which frame timestamps are represented. For fixed-fps content,
        * timebase should be 1/framerate and timestamp increments should be
        * identical to 1. */
    ost->st->time_base = {1, STREAM_FRAME_RATE};
    c->time_base       = ost->st->time_base;
    c->gop_size = 12; /* emit one intra frame every twelve frames at most */

    if (nvenc && USE_D3D11) {
        const std::string hw_device_name = "d3d11va";
        AVHWDeviceType    device_type    = av_hwdevice_find_type_by_name(hw_device_name.c_str());

        // set up hw device context
        AVBufferRef *hw_device_ctx;
        // const char*  device = "0"; // Default GPU (may be integrated in the case of switchable graphics!)
        const char*  device = "1";
        ret = av_hwdevice_ctx_create(&hw_device_ctx, device_type, device, nullptr, 0);

        if (ret < 0) {
            fprintf(stderr, "Could not create hwdevice context; %s", av_err2str(ret));
        }

        set_hwframe_ctx(c, hw_device_ctx);
        c->pix_fmt = AV_PIX_FMT_D3D11;
    } else if (nvenc && !USE_D3D11)
        c->pix_fmt = AV_PIX_FMT_BGR0;
    else
        c->pix_fmt = STREAM_PIX_FMT;

    if (c->codec_id == AV_CODEC_ID_MPEG2VIDEO) {
        /* just for testing, we also add B-frames */
        c->max_b_frames = 2;
    }

    if (c->codec_id == AV_CODEC_ID_MPEG1VIDEO) {
        /* Needed to avoid using macroblocks in which some coeffs overflow.
            * This does not happen with normal video, it just happens here as
            * the motion of the chroma plane does not match the luma plane. */
        c->mb_decision = 2;
    }

    /* Some formats want stream headers to be separate. */
    if (oc->oformat->flags & AVFMT_GLOBALHEADER)
        c->flags |= AV_CODEC_FLAG_GLOBAL_HEADER;
}


  


• can ffmpeg detect monochromatic parts of a video and conditionally set filter parameters accordingly

  15 juillet 2021, par Silentfury

  I have input videos that switch back and forth between coloured and monochrome/grey parts (frames that only contain luma information). Is there a way in ffmpeg to detect whether the current frame is coloured or monochrome and is it possible to use this information to change the parameters of a filter ?

  


  As an example I want to apply the mpdecimate filter and I want it to use a "hi" parameter of 50 (per pixel) for coloured frames and 20 for monochrome frames.

  


  So in pseudo code I want something something like

  


  -vf mpdecimate=hi=if(coloured,64*50,64*20)

  


  but obviously the "coloured" condition doesnt work that way.

  


  Is that somehow achievable ?

  


  PS :
Concernig the detection of monochrome frames I have identified two filters that have a related functionality but I am unable to use the information in another filter :

  


  

  • "geq" can get a good idea of frame colors using cbsum(x,y) and/or crsum(x,y)

  


  • "signalstats" can use the property "SATMAX" to tell about the color level of a frame

  


  


  Is there any way to use the resulting number of any of the two above filters in another filter ?

  


• avcodec/bsf : switch to av_get_token to parse bsf list string

  3 juillet 2021, par Gyan Doshi

  avcodec/bsf : switch to av_get_token to parse bsf list string
  The recently added setts bsf makes use of the eval API whose
  
  expressions can contain commas. The existing parsing in
  
  av_bsf_list_parse_str() uses av_strtok to naively split
  
  the string at commas, thus preventing the use of setts filter
  
  with expressions containing commas.
  av_get_token can work with escaped commas, allowing full use of setts.
  

  • [DH] libavcodec/bsf.c