Can you pack and reuse pointers? Trying to make LibRealsense garbage free

[btshrewsbury]

Hi Saudet!

I am trying to get depth data from a few real sense cameras into a real time java thread and I would like to avoid calling the GC if possible. I was able to profile and it looks like most of the allocations on the Java side are related to new pointer allocations (Pointer, rs2_frame, and rs2_stream_profile). Do you know if it's possible to reuse pointers instead of creating new ones each javacpp call? I think we can manually call the deallocation of native memory within the realtime thread once we are done with the byte buffers. I don't fully understand how the pointers get allocated and managed under the hood, so I am not really sure where to go from here.

Javacpp is very helpful, Thank you!

[saudet]

If you're talking about deterministic deallocation, we can do that with Pointer.close(), but PointerScope is more convenient:
http://bytedeco.org/news/2018/07/17/bytedeco-as-distribution/
If you're talking about reusing native memory for RealSense, that's going to depend on its API and what it's capable of.
Or are you talking about something else?

[btshrewsbury]

I may not be describing it well so code might just be easier? I just don't want to call new. It looks like polling allows us to pack rs2_frame without allocating. rs2_get_frame_data requires a new BytePointer though. I just want to reuse it every tick. I was looking at manually calling deallocate, null, then init on the packablePointer, but I haven't tried it yet.

rs2_frame depth_frame = new rs2_frame();
BytePointer packablePointer= new BytePointer();

while (true) 
{
     realsense2.rs2_pipeline_poll_for_frames(pipe, depth_frame, e);

     //this is the way it's typically 
     BytePointer newInstantiatedPointer = new BytePointer(realsense2.rs2_get_frame_data(depth_frame, e)); done

      //this is what I want to do, so I only have to instantiate the byte pointer once. 
     packablePointer.set(realsense2.rs2_get_frame_data(depth_frame, e));

     //do stuff with the packed pointer data

     //allow the C side to release the data
     packablePointer.deallocate();
}

[saudet]

I've added an rs2_get_frame_data_address() overload that doesn't create new objects and returns the address as a long in commit bytedeco/javacpp-presets@7f0122b. You can use that value to set the address of, for example, your MutableBytePointer. Please give it a try with the snapshots: http://bytedeco.org/builds/

[btshrewsbury]

It works great, Thank you!

Yourkit is complaining about depth_pointer.get(depth_bytes, 0, depth_data_size) now. It looks like just about every pointer method returns a new pointer so you can perform nice method chaining. I tried manually calling get() while incrementing the offset to go through the bytes manually, but that calls a method under the hood that also creates a pointer. Is there a way to get the contents of a byte array without creating a new pointer? Sorry to be a pain in the ass.

I tried to use asBuffer(), but the result ByteBuffer has a limit of 1 and a capacity of 1.

[saudet]

You mean this method? That doesn't create any new objects, no.
http://bytedeco.org/javacpp/apidocs/org/bytedeco/javacpp/BytePointer.html#get-byte:A-int-int-

What you want to use though is probably an indexer:
http://bytedeco.org/news/2014/12/23/third-release/

[btshrewsbury]

Using JVM statistics the memory usage stays constant so everything is working. No idea what Yourkit thinks is going on. Thank you again for the help! I will look at the indexer to help speed up moving the data around.

[saudet]

Never fully trust what a profiler tells you :)