sdk link issue with libnvinfer.so rpath and libcudla.so

[aurelien-enchanted-tools]

Hello, I met an issue with the sdk and cmake during the link step

poky: langdale
meta-tegra: langdale

ld: warning: libcudla.so.1, needed by /home/yoctouser/sdk/sysroots/armv8a-custom-linux/usr/lib/libnvinfer.so, not found (try using -rpath or -rpath-link)
ld: /home/yoctouser/sdk/sysroots/armv8a-custom-linux/usr/lib/libnvinfer.so: undefined reference to `[email protected]'

In the sdk, I have got:

• /usr/local/oe-sdk-hardcoded-buildpath/sysroots/armv8a-custom-linux/usr/lib/libnvinfer.so and
• /usr/local/oe-sdk-hardcoded-buildpath/sysroots/armv8a-custom-linux/usr/local/cuda-11.4/lib/libcudla.so.1

I check the c++ method cudlaSubmitTask is in /usr/local/oe-sdk-hardcoded-buildpath/sysroots/armv8a-custom-linux/usr/local/cuda-11.4/lib/libcudla.so.1

In the image:

TOOLCHAIN_HOST_TASK += " \
    nativesdk-packagegroup-cuda-sdk-host \
"

I would like to know:

• how to include the /usr/local/oe-sdk-hardcoded-buildpath/sysroots/armv8a-custom-linux/usr/local/cuda-11.4/lib/ in the libnvinfer's RPATH, or better
• how to fix the issue directly in the tensorrt-core recipe which populates the libnvinfer.so

Thank you in advance.

[madisongh]

The necessary compiler and linker flags/options to make the CUDA libraries available should get added automatically by CMake if you've told it you're using CUDA, either by specifying CUDA in the LANGUAGES clause in your project() command, or through enable_language(CUDA) (or, if you're using a very old version of CMake, by using FindCUDA()). The SDK sets up environment variables and provides a CMake "toolchain" file for this purpose. But you have to tell CMake about CUDA being used in your project.

[aurelien-enchanted-tools]

langdale cmake version: 3.24.2

I added in my yocto image config

TOOLCHAIN_HOST_TASK += " \
   nativesdk-cuda-environment \
   nativesdk-packagegroup-cuda-sdk-host \
"
 
TOOLCHAIN_TARGET_TASK += "\
    cuda-target-environment \
"

and in my CMakeLists.txt build with the sdk

project(custom LANGUAGES C CXX CUDA)
find_package(CUDA REQUIRED)

and I obtain the same error as previous.

[madisongh]

Can you enable tracing (use --trace-expand on the cmake command) and see where the additions might be getting lost? I do have folks with cmake projects that call on the TensorRT libraries that don't run into this issue, so I'm wondering if either there's something in your environment that's causing the CUDA additions to get bypassed, or there's something in your cmake files that might be overriding what the macros are inserting for you into the compilation and linking flags.