rtpL16depay, rtpL24depay: support depayloading in place
Submitted by Petr Kulhavy
Link to original bug (#773833)
Description
It seems that rtpL16depay, rtpL24depay (and I guess also other rtp-depay functions) are failing to depayload in place. It always allocates a new buffer and copies the payload.
This is the log I'm seeing:
0:00:13.489419481 1060 0x1ec5150 DEBUG rtpL16depay gstrtpL16depay.c:243:gst_rtp_L16_depay_process:<rtpl16depay0> got payload of 192 bytes
0:00:13.489596481 1060 0x1ec5150 LOG GST_BUFFER gstbuffer.c:798:gst_buffer_new: new 0x75c05370
0:00:13.489760981 1060 0x1ec5150 LOG GST_BUFFER gstbuffer.c:2038:gst_buffer_copy_region: new region copy 0x75c05370 of 0x75c054b0 12-192
0:00:13.490969772 1060 0x1ec5150 LOG GST_BUFFER gstbuffer.c:514:gst_buffer_copy_into: copy 0x75c054b0 to 0x75c05370, offset 12-192/204
0:00:13.493352979 1060 0x1ec5150 TRACE GST_LOCKING gstminiobject.c:177:gst_mini_object_lock: lock 0x1ecb690: state 00020000, access_mode 3
0:00:13.493546312 1060 0x1ec5150 DEBUG GST_LOCKING gstminiobject.c:212:gst_mini_object_lock: lock failed 0x1ecb690: state 00020000, access_mode 3
0:00:13.493717020 1060 0x1ec5150 DEBUG GST_MEMORY gstmemory.c:317:gst_memory_map: mem 0x1ecb690: lock 3 failed
0:00:13.493910270 1060 0x1ec5150 TRACE GST_REFCOUNTING gstobject.c:249:gst_object_ref:<allocatorsysmem0> 0x1ddc010 ref 6->7
0:00:13.494091270 1060 0x1ec5150 DEBUG GST_MEMORY gstmemory.c:138:gst_memory_init: new memory 0x1ecb578, maxsize:195 offset:0 size:192
0:00:13.494364978 1060 0x1ec5150 DEBUG GST_PERFORMANCE gstallocator.c:462:_sysmem_copy: memcpy 192 memory 0x1ecb690 -> 0x1ecb578
And this is the command:
GST_DEBUG=9 gst-launch-1.0 audiotestsrc is-live=true wave=silence samplesperbuffer=48 ! "audio/x-raw,format=(string)S16BE,rate=(int)48000,channels=(int)2,channel-mask=(int)0x3" ! rtpL16pay ! rtpL16depay ! fakesink
2>&1 | grep -B 5 copy
It happens in gst_audio_buffer_reorder_channels() when mapping the buffer read-write. It doesn't really matter what the source is, the same happens with udpsrc.
--
The other thing here is that if the channels don't require reordering (like in my case) the buffer shouldn't be remapped read-write. Otherwise the potential buffer copy is redundant.