Open GoogleCodeExporter opened 8 years ago
At my employer we too have to use the file provider API and have made changes
to that code. While we do have actual files to read and write we also need to
build an MP4 in memory. Specifically, we build MP4 files of short coded video
sequences that are guaranteed to begin with an I frame.
The issue with the file provider is that it does not provide an abstraction for
getting the file size. It assumes the construct is a real file and calls a
function that expects a genuine file descriptor. This may be OK in *NIX since
most things are file descriptors even if it is not a file on disk. However, in
Windows the API call to get file information fails because the file provider is
not a real file (that is, the file provider cannot return a handle that is
compatible with the Win32 API call for getting file size).
As stated, I have made changes to accommodate my needs. I don't want to
maintain a personal vendor branch and would like to contribute my changes back
to the community. I have provided a patch file of my changes. Allow me to
describe those changes. I'll be as brief as possible.
I strived for changes that are non-breaking to existing compiled projects.
However, this was not 100% possible. However, reading the discussion from the
OP's link shows that a breaking change isn't entirely taboo.
The first change was to provide additional MP4Create functions. I followed the
style of the MP4Read functions. Thus, I have added two new functions named
MP4CreatProvider and MP4CreateProviderEx. This is a non-breaking change. The
implementation of MP4Create and MP4CreateEx have been modified to simply call
into MP4CreateProviderEx. Default arguments have been preserved. The original
MP4Create and MP4CreateEx functions retain their behavior in that a NULL
provider is used under the hood, which results in the StandardFileProvider
instance being used.
The second change was to add an additional function pointer to the
MP4FileProvider interface. This is a breaking change. I had a choice between
adding a "tell" operation or a "getSize" operation. With "tell" you would seek
to the end and then call "tell" to get the current position. A problem of that
strategy is it could potentially be expensive, say you were dealing with a
large MP4 movie from a SAN drive (maybe a unrealistic example but you get my
point). Thus, I decided on "getSize" because that implies a cheap operation
(reading the file meta data as opposed to seeking through the whole file).
Lastly, for all intents and purposes I made a very conscience effort to keep
the style and design of the existing code. All spacing, alignments, naming
style, parameter placement, comments, etc., have been written to faithfully
match its neighboring code. I hope this helps with the adoption of the patch.
Please also note that I develop for the Windows platform so my changes are
tested on that platform. I do not have as thorough of a test for *NIX
platforms.
Any feedback is appreciated. Thank you.
Nicholas
Original comment by Nicholas...@gmail.com
on 18 Aug 2014 at 6:47
Attachments:
I think it's fine to make breaking changes to this API, because I agree the
design is more or less broken (primarily due to the file size issue you
mention).
Original comment by kid...@gmail.com
on 14 Oct 2014 at 3:26
I have committed my patch. My patch does not represent a major redesign of the
feature. It only addresses a single point of failure, namely the need of the
provider API to allow querying the size of the file.
As a Windows developer the code changes have been through much use on the
Windows platform. For *NIX platforms I fired up an openSUSE Linux VM. I have
built the library with my patch applied. There was a compiler error in my
patch and has been fixed. I created a test harness app that reads a raw H.264
file (from the encoder cards that I work with at my job) and generates an MP4
file using the custom file provider that adapts an in-memory buffer as a file
(note: this is my use case that is the motivation for this patch, this is the
work flow that I have to deal with at my job). The test harness dumped the
memory buffer to file when it was finished. I was able to play the file in
VLC. My H.264 analysis programs were also able to consume the file.
Original comment by Nicholas...@gmail.com
on 17 Oct 2014 at 10:47
Original issue reported on code.google.com by
Andree.e...@gmail.com
on 4 Apr 2014 at 12:37