nvmmappsrc

Zero-copy NVMM IPC source. Connects to an nvmmsink (or compatible producer) over its unix socket, receives the pool buffers' NvBufSurfaceMapParams + DMA-buf fds, imports them with NvBufSurfaceImport, and pushes NVMM GstBuffers downstream — reading the producer's GPU memory in place, no copy. See Zero-copy IPC.

Src caps: video/x-raw(memory:NVMM), {NV12, RGBA, I420, BGRA}.

Properties

Property	Type	Default	Description
shm-name	string	/nvmm_sink_0	POSIX shared-memory segment to read from
is-live	bool	true	Whether this source is live

Example (consumer)

# NVMM -> system memory for JPEG (VIC via nvvidconv, not CPU videoconvert)
gst-launch-1.0 -e nvmmappsrc shm-name=/ipc_test is-live=true ! \
  'video/x-raw(memory:NVMM)' ! nvvidconv ! 'video/x-raw,format=I420' ! \
  nvjpegenc ! filesink location=ipc_out.jpg

# True zero-copy consumer: hardware encoder reads NVMM directly (Xavier — Orin NX has no NVENC)
gst-launch-1.0 -e nvmmappsrc shm-name=/ipc_test is-live=true ! \
  'video/x-raw(memory:NVMM),format=NV12' ! nvv4l2h264enc ! h264parse ! qtmux ! filesink location=out.mp4

nvmmappsrc emits memory:NVMM. A CPU element (videoconvert) can't consume it — go through nvvidconv (VIC), or feed a hardware consumer like nvv4l2h264enc that takes NVMM directly.

DeepStream interop

nvmmappsrc pushes NvBufSurface-backed NVMM buffers, which nvstreammux accepts directly. So one producer published once (by nvmmsink) can fan out to multiple independent DeepStream inference processes, each importing the pool fds in place — no extra GPU copy per consumer. Feed nvstreammux (batch the single stream), then nvinfer:

Verified on Orin

Tested on Orin NX (JP6) with DeepStream 7.1: an nvmmsink producer feeding nvmmappsrc ! nvstreammux ! nvinfer ran to EOS with inference on the imported buffers. Adjust nvstreammux properties and the nvinfer config-file-path to your install; set nvinfer batch-size=1 for a single stream (the stock config defaults to a batch-30 INT8 engine).

# Zero-copy IPC source -> DeepStream detector
gst-launch-1.0 -e \
  nvmmappsrc shm-name=/ipc_test do-timestamp=true is-live=true \
  ! 'video/x-raw(memory:NVMM),format=NV12' \
  ! nvstreammux batch-size=1 width=1920 height=1080 batched-push-timeout=40000 \
  ! nvinfer config-file-path=pgie.txt batch-size=1 \
  ! nvvideoconvert ! nvdsosd ! nv3dsink

Only pixels cross the IPC boundary — not DeepStream metadata

NvDsBatchMeta and per-object metadata (bounding boxes, classes, tracker IDs) are attached to the GstBuffer, not to the NvBufSurface. nvmmsink / nvmmappsrc transfer only the surface, so metadata produced by an nvinfer in the producer process does not arrive here. Run inference in the consumer (as above), or — if you only need the visualized result — burn the overlay into the pixels with nvdsosd before nvmmsink in the producer.