nvmmsink

Shares NVMM video frames across processes via a GPU-copy pool. Each incoming frame is copied GPU-to-GPU into a fixed pool of NVMM buffers (via NvBufSurfTransform on the VIC, which also de-tiles BLOCK_LINEAR input into the pool's PITCH_LINEAR layout), and the pool buffers' DMA-buf fds are handed to consumers over a unix-domain socket (SCM_RIGHTS). Consumers import the fds and read GPU memory directly. See Zero-copy IPC.

Sink 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 name (e.g. /cam1)
pool-size	int (13–16)	16	Number of NVMM buffers in the shared pool

Why the minimum is 13

nvmmappsrc holds a reference on up to RELEASE_DELAY (12) in-flight buffers (to cover the consumer's hardware-encoder pipeline depth). A pool of ≤ 12 lets a steady consumer hold every slot and starve the producer (the stream stalls after ~pool-size frames). The property therefore enforces a minimum of 13 (RELEASE_DELAY + 1); the default is 16.

Example (producer)

gst-launch-1.0 videotestsrc num-buffers=300 pattern=smpte ! \
  'video/x-raw,width=640,height=480,format=I420,framerate=10/1' ! \
  nvvidconv ! 'video/x-raw(memory:NVMM),format=NV12' ! \
  nvmmsink shm-name=/ipc_test sync=true

The producer makes one GPU-side copy into the shared pool; consumers then import without further copies (single-copy producer, zero-copy consumers).

DeepStream interop

nvmmsink takes video/x-raw(memory:NVMM) (NV12 / RGBA) — the same surfaces a DeepStream pipeline carries — so you can tap a DeepStream pipeline and publish its frames to other processes. Insert it on an NVMM branch after the stage you want to share:

Verified on Orin

Tested on Orin NX (JP6) with DeepStream 7.1: this pipeline published annotated NVMM frames through nvmmsink, and a separate nvmmappsrc consumer imported and re-encoded them. Set nvinfer batch-size=1 for a single stream (the stock config defaults to a batch-30 INT8 engine); adjust config-file-path to your install.

# Run inference, burn the overlay into the pixels, publish annotated NVMM frames
gst-launch-1.0 \
  filesrc location=video.mp4 ! qtdemux ! h264parse ! nvv4l2decoder \
  ! 'video/x-raw(memory:NVMM)' \
  ! nvstreammux batch-size=1 width=1920 height=1080 \
  ! nvinfer config-file-path=pgie.txt batch-size=1 \
  ! nvvideoconvert ! nvdsosd \
  ! nvvideoconvert ! 'video/x-raw(memory:NVMM),format=NV12' \
  ! nvmmsink shm-name=/ds_annotated

Only pixels cross the IPC boundary — not DeepStream metadata

NvDsBatchMeta and per-object metadata (boxes, classes, tracker IDs) live on the GstBuffer, not the NvBufSurface. nvmmsink ships only the surface, so downstream consumers receive the frames but not the inference results as metadata. To share the result, render it into the pixels with nvdsosd before nvmmsink (as above). Consumers then read finished, annotated frames; they cannot recover the structured detections.