[Issue]: rocJPEG tests fail on CI due to VA-API driver mismatch

[daniellegillai]

Problem Description

CTests for rocJPEG (basic_test_rocjpeg_gray and basic_test_rocjpeg_rgb) fail due to a mismatch between the libva2 version (VA-API 1.20) and Mesa driver in ROCm (VA-API 1.22).

rocJPEG links to the libva 1.20 runtime. It will try to dynamically load the driver at /opt/rocm/lib/rocm_sysdeps/lib/radeonsi_drv_video.so. This leads to the following libva error:
libva error: /opt/rocm/lib/rocm_sysdeps/lib/radeonsi_drv_video.so has no function __vaDriverInit_1_0.
nm shows that this driver exports __vaDriverInit_1_22.

Meanwhile, when instead rocJPEG is directed to use the driver at /usr/lib/x86_64-linux-gnu/libgallium-*.so from ffmpeg, the tests pass without an error. nm shows that this driver exports __vaDriverInit_1_20.

See "Produce more debug info" for more detail.

Operating System

Ubuntu 24.04.4 LTS (Noble Numbat)

CPU

AMD Ryzen 9 9950X 16-Core Processor

GPU

AMD Radeon RX 9070 XT

ROCm Version

7.15.0

ROCm Component

No response

Steps to Reproduce

Setup

Use the same docker image as the CI:

sudo docker run -it --rm --device /dev/kfd --device /dev/dri mivisionx/ubuntu-24.04:rocm-20260710 bash

In the container, run:

export ROCM_PATH=/opt/rocm
export LD_LIBRARY_PATH=/opt/rocm/lib
export LIBVA_DRIVERS_PATH=/opt/rocm/lib/rocm_sysdeps/lib
git clone https://github.com/rocm/rocAL.git
cd  rocAL

(Still in the container) Run the same library / dependency set up as in the CI:

apt-get update -qq
apt-get install -y --no-install-recommends build-essential cmake git python3 \
libprotobuf-dev libprotoc-dev protobuf-compiler libturbojpeg0-dev liblmdb-dev libjpeg-dev \
pkg-config ffmpeg libavcodec-dev libavformat-dev libavutil-dev libswscale-dev
git clone --depth 1 https://github.com/Tencent/rapidjson.git
cd rapidjson
mkdir build
cd build
cmake ../ -DCMAKE_POLICY_VERSION_MINIMUM=3.5 -DRAPIDJSON_BUILD_EXAMPLES=OFF -DRAPIDJSON_BUILD_TESTS=OFF -DRAPIDJSON_BUILD_DOC=OFF
make -j$(nproc)
make install
cd ../..
git clone --depth 1 --filter=blob:none --sparse https://github.com/ROCm/rocm-libraries.git
cd rocm-libraries
git sparse-checkout set projects/rpp
cd projects/rpp
mkdir build
cd build
cmake ..
make -j$(nproc)
make install
cd ../../../../
git clone --depth 1 https://github.com/ROCm/MIVisionX.git
cd MIVisionX
mkdir build && cd build
cmake .. \
  -DBACKEND=HIP \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_C_COMPILER="${ROCM_PATH}/bin/amdclang" \
  -DCMAKE_CXX_COMPILER="${ROCM_PATH}/bin/amdclang++"
make -j$(nproc)
make install
cd ../..
mkdir build
cd build
cmake .. -DBUILD_PYPACKAGE=OFF -DCMAKE_DISABLE_FIND_PACKAGE_rocdecode=ON
make -j$(nproc)
make install

Produce the error

Run ctest on basic_test_cpu to build the test, then run it on rocjpeg.

cd /rocAL/build
ctest -R basic_test_cpu
cd tests/cpp_api/basic_test

./basic_test ${ROCM_PATH}/share/rocal/test/data/images/AMD-tinyDataSet ${ROCM_PATH}/share/rocal/test/data/images/AMD-tinyDataSet-val.txt 1 1 224 224 0 1 1 # basic_test_rocjpeg_gray
./basic_test ${ROCM_PATH}/share/rocal/test/data/images/AMD-tinyDataSet ${ROCM_PATH}/share/rocal/test/data/images/AMD-tinyDataSet-val.txt 1 1 224 224 1 1 1 # basic_test_rocjpeg_rgb

Produce more debug info

Command: LIBVA_MESSAGING_LEVEL=1 ./basic_test ${ROCM_PATH}/share/rocal/test/data/images/AMD-tinyDataSet ${ROCM_PATH}/share/rocal/test/data/images/AMD-tinyDataSet-val.txt 1 1 224 224 1 1 1 2>&1 | grep -i driver
Output: libva error: /opt/rocm/lib/rocm_sysdeps/lib/radeonsi_drv_video.so has no function __vaDriverInit_1_0

Command: nm -D /opt/rocm/lib/rocm_sysdeps/lib/radeonsi_drv_video.so | grep vaDriver
Output: 0000000000028fb0 T __vaDriverInit_1_22@@AMDROCM_SYSDEPS_1.0

Command: nm -D /usr/lib/x86_64-linux-gnu/libgallium-25.2.8-0ubuntu0.24.04.2.so | grep vaDriver
Output: 0000000000100a80 T __vaDriverInit_1_20@@libgallium-25.2.8-0ubuntu0.24.04.2.so

Command:

apt install -y vainfo
vainfo 2>&1 | grep version

Output: libva info: VA-API version 1.20.0

Command: apt-cache policy libva2
Output:

libva2:
  Installed: 2.20.0-2ubuntu0.2
  Candidate: 2.20.0-2ubuntu0.2
  Version table:
 *** 2.20.0-2ubuntu0.2 500
        500 http://archive.ubuntu.com/ubuntu noble-updates/universe amd64 Packages
        500 http://security.ubuntu.com/ubuntu noble-security/universe amd64 Packages
        100 /var/lib/dpkg/status
     2.20.0-2build1 500
        500 http://archive.ubuntu.com/ubuntu noble/universe amd64 Packages

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

rocminfo --support output

Paste output here

Additional Information

No response

[AryanSalmanpour]

@kiritigowda @daniellegillai Please follow these instructions before running the rocJPEG CTest. If multiple versions of libva are installed, use LD_PRELOAD to load the libva library from the TheRock path.

export ROCM_PATH=<path_to_the_ROCM_installation>

export LD_PRELOAD=$ROCM_PATH/lib/rocm_sysdeps/lib/librocm_sysdeps_va.so.2:$ROCM_PATH/lib/rocm_sysdeps/lib/librocm_sysdeps_va-drm.so.2

export LIBVA_DRIVERS_PATH=$ROCM_PATH/lib/rocm_sysdeps/lib/

[daniellegillai]

@AryanSalmanpour Adding the LD_PRELOAD flag export fixed it locally, thanks! Will update the workflow in #482 to add that.