tests

Hardware regression tests

Run on the supported Iris board. These are hardware integration tests, not host unit tests. C/C++ files include build commands in their headers; FFmpeg tests also require ffprobe and software H.264/HEVC decoding.

meson test -C build also checks CAPTURE buffer budgeting without hardware.

Test	Coverage
decoder/main10.cpp	HEVC Main10 / AV1 8/10-bit capabilities, config validation, surface enumeration, and context bit-depth selection.
decoder/av1.py	FFmpeg/GStreamer AV1 pixels against dav1d: 720p–4K, 8/10-bit, hidden references, tiles, lossless, still pictures, super-resolution and global motion.
decoder/prime-import.cpp	PRIME1/PRIME2 NV12/P010 imports, malformed descriptors, shared storage, and caller FD ownership.
decoder/error-recovery.py	Corrupt HEVC, queue errors, neutral NV12/P010 pixels, B frames, consecutive errors, DRM mapping, and persistent/imported storage.
encoder/ffmpeg-smoke.py	Profiles, rate control, dimensions, frame counts, keyframes, and decoded quality.
encoder/api-smoke.cpp	Uploads, direct/staged input, rejected requests, coded buffers, and resource cleanup.
encoder/async.cpp	Queued versus serial output, QP/IDR changes, non-IDR I requests, synchronization, timeouts, and pending resource destruction.
encoder/writable-export.cpp	GPU-written input, stable DMA-BUF exports, and GPU fence synchronization across non-IDR GOPs.

To select an uninstalled driver before running a test:

export LIBVA_DRIVER_NAME=qcom LIBVA_DRIVERS_PATH="$PWD/build"

The PRIME import test requires /dev/dma_heap/system. It checks padded layouts and nonzero offsets, and verifies that importing and exporting retain the same DMA-BUF rather than copying its pixels.

Run AV1 on a board exposing the V4L2 AV1 format, with the NOSHOW kernel patch: python3 tests/decoder/av1.py tmp/av1 --driver-dir build --gstreamer --extended. Requires FFmpeg with libaom/dav1d; --gstreamer also needs the va plugin and GStreamer app/video development libraries. --extended adds coding-tool cases and requires aomenc. Passing decoded files are removed; logs and streams remain.

Run error recovery with python3 tests/decoder/error-recovery.py tmp/recovery --driver-dir build on a kernel with the error-timestamp patch. It generates HEVC samples with libx265 and builds the adjacent preload shims. Besides real corruption, it checks that injected errors replace only the affected frames, exports retain their DMA-BUF, and failed fills or missing timestamps do not report successful recovery.

The FFmpeg, async, and writable-export tests accept --hevc to exercise HEVC Main instead of H.264. The HEVC async test also checks rejected picture types, tiles, partial slices, invalid QP/filter values, and long-term references.

Place generated streams and logs under tmp/. The async and writable-export programs take an existing output directory as their first argument.

Optional failure injection helpers (build commands are in their headers):

• reject-coherent.c: load with LD_PRELOAD to test the CPU/MMAP fallback on hardware that normally supports coherent GEM. Works with all encoder tests.
• delay-completion.c: load with LD_PRELOAD and set IRIS_TEST_DELAY_COMPLETION=1 when running async to test caller timeouts and the hardware wait watchdog.
• decoder/error-completion.c: load with LD_PRELOAD and set IRIS_TEST_ERROR_QUEUE=capture, output, or both to inject errors at decode token 46. IRIS_TEST_ERROR_FIRST and IRIS_TEST_ERROR_COUNT select another frame or consecutive errors. Errors must complete without a sync or context destruction timeout, including late CAPTURE after OUTPUT errors.

For staged async input, set IRIS_TEST_STAGING=1; add IRIS_VAAPI_COPY=cpu to exercise CPU staging. Without that override, staging uses the build's default copy path.

Build options

python3 tests/build/gpu-copy.py tmp/build-options runs without Iris hardware. Install all GPU-copy development dependencies first. It masks each dependency in turn and checks auto, enabled, and disabled, source selection, linked libraries, and driver loading. Logs and builds remain in the output directory.

On the board, also run the decoder recovery and encoder staging tests against -Dgpu_copy=disabled without setting IRIS_VAAPI_COPY, to exercise the default CPU path. Explicit IRIS_VAAPI_COPY=gpu must fail only when a copy is needed; DMA-BUF zero-copy remains usable.

CDSP VPP

meson test -C build --print-errorlogs checks MSM allocation cache modes, skeleton selection and HVX worker limits without hardware. Build with CDSP enabled, then run on the target (upstream FastRPC kernel ABI):

c++ -O2 -std=c++17 -rdynamic tests/vpp/scaling.cpp -o build/vpp-scaling \
  $(pkg-config --cflags --libs libva libva-drm libdrm vulkan) -ldl \
  -Lbuild/src/cdsp -liris_vpp
export LIBVA_DRIVER_NAME=qcom LIBVA_DRIVERS_PATH="$PWD/build"
export LD_LIBRARY_PATH="$PWD/build/src/cdsp${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
export ADSP_LIBRARY_PATH="$PWD/build/src/cdsp${ADSP_LIBRARY_PATH:+;$ADSP_LIBRARY_PATH}"
./build/vpp-scaling

Checks NV12/P010 scaling, conversion and rotation, rounding, crop/output layouts, DMA-BUF identity, persistent mapping, context/buffer lifetime and rotation with five live contexts sharing scratch resources. Also run against a -Dfastrpc=bundled build. Pass vulkan to run the same VAAPI cases and check persistent Vulkan imports instead of FastRPC mappings. For a build without CDSP, add -DVPP_NO_CDSP and omit -Lbuild/src/cdsp -liris_vpp. --bench measures 100 synchronous submissions after a one-second warmup, excluding uploads and reference generation.

vpp/tonemap.cpp checks Vulkan HDR10 → SDR against a double-precision reference, including metadata changes, padded/cropped layouts, DMA-BUF identity, persistent imports and context lifetime. Build with the command in its header and run with the same driver environment; --geometry also checks combined HDR, scaling and rotation with the selected VPP backend. It reports warmed VAAPI call times for 1080p and 4K.

vpp/overlay.cpp checks both overlay backends against a double-precision reference, including RGB upload/PRIME round trips, clipping, scaling, alpha, multiple layers, rejected submissions and parameter lifetimes. Build with the command in its header. --geometry adds combined base scaling; --bench measures warmed VAAPI submissions including the base copy, excluding uploads and encoding. Pass cdsp or vulkan to test a single backend.

Optional ABI audit against SDK 6.7 (all six architectures):

python3 tests/vpp/check-abi.py /path/to/19.0.02/Tools /path/to/HexagonSDK6.7