Correctly Allocate And Fill Frame In FFmpeg

问题:

I am filling a Frame with a BGR image for encoding, and I am getting a memory leak. I think I got to the source of the problem but it appears to be a library issue instead. Since FFmpeg is such a mature library, I think I am misusing it and I would like to be instructed on how to do it correctly.

I am allocating a Frame using:

AVFrame *bgrFrame = av_frame_alloc();

And later I allocate the image in the Frame using:

av_image_alloc(bgrFrame->data, bgrFrame->linesize, bgrFrame->width, bgrFrame->height, AV_PIX_FMT_BGR24, 32);

Then I fill the image allocated using:

av_image_fill_pointers(bgrFrame->data, AV_PIX_FMT_BGR24, bgrFrame->height, originalBGRImage.data, bgrFrame->linesize);

Where originalBGRImage is an OpenCV Mat.

And this has a memory leak, apparently, av_image_alloc() allocates memory, and av_image_fill_pointers() also allocates memory, on the same pointers (I can see bgrFrame->data[0] changing between calls).

If I call

av_freep(&bgrFrame->data[0]);

After av_image_alloc(), it's fine, but if I call it after av_image_fill_pointers(), the program crashes, even though bgrFrame->data[0] is not NULL, which I find very curious.

Looking FFmpeg's av_image_alloc() source code, I see it calls av_image_fill_pointers() twice inside it, once allocating a buffer buff....and later in av_image_fill_pointers() source code, data[0] is substituted by the image pointer, which is (I think) the source of the memory leak, since data[0] was holding buf from the previous av_image_alloc() call.

So this brings the final question: What's the correct way of filling a frame with an image?.

回答1:

You should allocate your frame once.

AVFrame* alloc_picture(enum PixelFormat pix_fmt, int width, int height)
{
AVFrame *f = avcodec_alloc_frame();
if (!f)
    return NULL;
int size = avpicture_get_size(pix_fmt, width, height);
uint8_t *buffer = (uint8_t *) av_malloc(size);
if (!buffer) {
    av_free(f);
    return NULL;
}
avpicture_fill((AVPicture *)f, buffer, pix_fmt, width, height);
return f;
}

Yes, the cast (AVPicture*) is allowed https://stackoverflow.com/a/20498359/2079934 . In subsequent frames, you can write into the this frame. Since your OpenCV raw data is BGR and you need RGB or YUV, you can use sws_scale. In my application, I mirror vertically:

struct SwsContext* convertCtx = sws_getContext(width, height, PIX_FMT_RGB24, c->width, c->height, c->pix_fmt, SWS_FAST_BILINEAR, NULL, NULL, NULL);
avpicture_fill(&pic_raw, (uint8_t*)pixelBuffer, PIX_FMT_RGB24, width, height);
// flip
pic_raw.data[0] += (height - 1) * pic_raw.linesize[0]; 
pic_raw.linesize[0] *= -1;
sws_scale(convertCtx, pic_raw.data, pic_raw.linesize, 0, height, f->data, f->linesize);
out_size = avcodec_encode_video(c, outputBuffer, outputBufferSize, f);

(You can adapt PIX_FMT_RGB24 to your needs and read from cv::Mat without copying data.)

回答2:

av_fill_arrays() does the job. It will fill the frame's data[] and linesizes[] but not reallocating any memory.