Linux and Wayland - build scripts are very broken (list of fixes to make it work)

[danilw]

Manual fixes required to make it work:

Modify:

1. NRISamples/External/NRIFramework/External/NRI - apply this - Fix detection of wayland-client.h NRI#123
2. NRISamples/External/NRIFramework/External/NRI - apply this - Fix Linux gcc14 build NRIFramework#11

Now you can build - but it wont work (under Wayland for more info - read #11 ).

Continue:

3. Turn off Wayland completely - in cmake-gui or in config cmake files (of NRISamples) set these to off:

4. NRISamples/External/NRIFramework/External/NRI/Source/VK/SwapChainVK.hpp - line 96 edit to:

    // Surface caps
    uint32_t textureNum = 3;

Yes this problem can point on - "something wrong with NRI framework" or detection of number of swapchain images - I have no idea - if you want debug it.

Now it will build and can work under xcb/x11 (in Wayland also work).
But in system with multiple GPU - it will select:

Wrong CPU-emulation or integrated GPU.

To select discrete Nvidia GPU:

This app NRISamples does not have way to select GPU so:

Create file nvidia_icd.json

{
    "file_format_version" : "1.0.0",
    "ICD": {
        "library_path": "/usr/lib64/libGLX_nvidia.so.0",
        "api_version" : "1.3.211"
    }
}

And run:

export VK_ICD_FILENAMES="nvidia_icd.json"
_Bin/Triangle

or for AMD

export VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/radeon_icd.i686.json:/usr/share/vulkan/icd.d/radeon_icd.x86_64.json

Youl see in output:

And it works.

[dzhdanNV]

Thanks! Very appreciate the structured report!

Fixed. Please, re-clone from scratch and test again:

1. Fix detection of wayland-client.h NRI#123 - integrated
2. Fix Linux gcc14 build NRIFramework#11 - partially manually integrated (the problem has been solved in another way on MathLib side)
3. Turn off Wayland completely - turned off in CMake file
4. VK swap chain now silently clamps user-provided SwapChainDesc::textureNum to the supported range
5. Multiple GPUs: modern NRI must put discrete adapters in the beginning of the list

[dzhdanNV]

Turn off Wayland completely - turned off in CMake file

Added NRIF_USE_WAYLAND which is OFF by default.

[danilw]

Tested - cloning repo git clone https://github.com/NVIDIA-RTX/NRISamples --recursive
and building - it works and no error messages

Multiple GPUs: modern NRI must put discrete adapters in the beginning of the list

I have 3 GPU listed for Vulkan: (not in order, order is random - Vulkan):

1. llvmpipe
2. integrated iGPU (amd)
3. discrete Nvidia

Every other Vulkan app/demo/sample even vkcube correctly - and works/run obviously:

> vkcube
Selected WSI platform: xcb
Selected GPU 0: NVIDIA GeForce RTX 4060, type: DiscreteGpu

> vkcube --wsi wayland
Selected GPU 0: NVIDIA GeForce RTX 4060, type: DiscreteGpu

This NRISamples for every sample:

> _Bin/Triangle
Creating borderless window (1920, 1080)
Loading...
nri::INFO (DeviceVK.hpp:151) - VK::llvmpipe (LLVM 19.1.7, 256 bits) - Supported instance extensions:

Without set of VK_ICD_FILENAMES - it wont select anything else.

llvmpipe is CPU type https://registry.khronos.org/vulkan/specs/latest/man/html/VkPhysicalDeviceType.html VK_PHYSICAL_DEVICE_TYPE_CPU

> vkcube --gpu_number 2
Selected WSI platform: xcb
Selected GPU 2: llvmpipe (LLVM 19.1.7, 256 bits), type: Cpu

[dzhdanNV]

Thanks for verifications! Yes, I have realized that it's not DISCRETE and not INTEGRATED. I will improve NRI code tomorrow (sorry, it's too late for me today).

P.S. Thanks a ton for valuable posts!

[dzhdanNV]

If you have a minute (but I don't insist) could you, please, close this issue and create another one in NRI repo for wrong classification of llvmpipe?

[danilw]

NVIDIA-RTX/NRI#124