SDK guide¶

Libraries¶

Target	Headers	Contents
`AMDVA::amdva_runtime`	`runtime.hpp`, `preprocess.hpp`, `telemetry.hpp`	VA/HIP import (`VaSession`, `ImportedSurface`), VPE operations, image packing
`AMDVA::amdva_inference`	`model.hpp`, `batching.hpp`, `inference.hpp`	`ModelInfo`, `ModelExecutor`, task post-processors, shared batch executor
`AMDVA::amdva_gst_support`	`tensor_meta.hpp`, `video_preprocess.hpp`, `tracking.hpp`, `overlay.hpp`, `publishing.hpp`, `gst_telemetry.hpp`	GstAnalytics interop, tracking, drawing, publication
`AMDVA::amdva_media`	`vui.hpp`, `rtsp_server.hpp`, `egl_import.hpp`, `copy_audit.hpp`	encode helpers, RTSP serving, EGL import
`AMDVA::amdva_streaming`	`streaming/*.hpp`	`RtspIngest`, `SegmentRecorder`, `Transcoder`, `Catalog` (recordings, recorded analysis, application tables), playback sessions, `PreviewSlot`
`AMDVA::amdva_monitor`	`resource_monitor.hpp`	CPU/GPU/memory/power and DRM engine load

Shared libraries carry the SONAME lib<name>.so.<major>.<minor>; 0.x minor releases make no ABI promise, so rebuild applications when the minor version changes.

Plugins (gstamdanalytics, gstgva, gstgvametapublish*) install to lib/gstreamer-1.0. Add that directory to GST_PLUGIN_PATH when you use a non-system prefix.

Build and install¶

source scripts/env.sh                 # rootless dependency environment
cmake -S . -B build -DAMDVA_ROCM_ROOT="$AMDVA_ROCM_ROOT" -DCMAKE_BUILD_TYPE=RelWithDebInfo
cmake --build build -j8
cmake --install build --prefix /path/to/amdva-sdk

Useful options: AMDVA_BUILD_PLUGINS, AMDVA_BUILD_GVA_COMPAT (vendored DL Streamer elements), AMDVA_BUILD_STREAMING, AMDVA_HARDWARE_TESTS, and for the applications under examples/: AMDVA_BUILD_EXAMPLES, AMDVA_BUILD_NVR, AMDVA_BUILD_NVR_GUI (needs Qt6), AMDVA_BUILD_DASHBOARD, AMDVA_BUILD_RTSP_FILES.

Optional dependencies, all ON by default: AMDVA_WITH_MQTT (Eclipse Paho), AMDVA_WITH_KAFKA (librdkafka) for result publication, and AMDVA_WITH_PYTHON (embedded CPython) for amdpython. A minimal framework build:

cmake -S . -B build-min -DAMDVA_ROCM_ROOT="$AMDVA_ROCM_ROOT" -DAMDVA_BUILD_EXAMPLES=OFF \
  -DAMDVA_WITH_MQTT=OFF -DAMDVA_WITH_KAFKA=OFF -DAMDVA_WITH_PYTHON=OFF

AMDVA_SYSROOT (default .deps/sysroot) locates the rootless development packages. The installed package provides amdva_use_host_glib() and amdva_use_host_sqlite() (AmdvaHostLibraries.cmake) so applications link the host GLib/SQLite runtime like the framework does.

Consuming from CMake¶

find_package(AMDVideoAnalytics REQUIRED)
add_executable(app main.cpp)
target_link_libraries(app PRIVATE AMDVA::amdva_media)

examples/sdk/CMakeLists.txt is a standalone project:

cmake -S examples/sdk -B build-sdk -DCMAKE_PREFIX_PATH=/path/to/amdva-sdk

Examples¶

Program / script	Shows
`examples/sdk/transcode.cpp`	decode, detect, overlay, encode to a file or RTSP
`examples/nvr/headless.cpp`	headless NVR service on the sample's controller (RTSP ingest, recording, AI)
`examples/nvr`, `examples/resource-dashboard`, `examples/rtsp-files`	complete applications; see each README
`examples/sdk/playback.cpp`, `timeline.cpp`, `keyframe-preview.cpp`	recorded-video APIs
`examples/sdk/telemetry.cpp`	decode and presentation timing
`examples/pipelines/*.sh`	`gst-launch` recipes: annotate to file, segment, multistream
`examples/metadata_callback.py`	`amdpython` callback

Tests¶

ctest --test-dir build                        # no GPU needed
cmake -S . -B build -DAMDVA_HARDWARE_TESTS=ON \
  -DAMDVA_DETECTION_MODELS=/dir/with/yolov8n-b1.mxr+model-info.json \
  -DAMDVA_TEST_VIDEO=/path/test.h264
ctest --test-dir build -R "encode|framework"  # needs the AMD GPU

Hardware tests include encode-contract (H.264/H.265/AV1, NV12 and BGRA input, BT.709 signalling) and encode-annotated-contract (decode, detect, overlay, encode through the DMA-BUF path).

Dependencies and licenses¶

scripts/bootstrap.py builds a locked GStreamer 1.28.2 with the GstAnalytics patch and a rootless sysroot under .deps/. DL Streamer sources are vendored in third_party/dlstreamer (MIT). See THIRD_PARTY_LICENSES.md.