Feature request: Shared fences on Vulkan for Profiling CudaInterop

[julcst]

I saw in the source code that shared fences are not functional for Vulkan:

NativeHandle Fence::getNativeHandle() const
{
    gfx::InteropHandle gfxNativeHandle = {};
    FALCOR_GFX_CALL(mGfxFence->getNativeHandle(&gfxNativeHandle));
#if FALCOR_HAS_D3D12
    if (mpDevice->getType() == Device::Type::D3D12)
        return NativeHandle(reinterpret_cast<ID3D12Fence*>(gfxNativeHandle.handleValue));
#endif
#if FALCOR_HAS_VULKAN
        // currently not supported
#endif
    return {};
}

Is this possible or planned? It is need to correctly profile Cuda kernels on Linux (tiny-cuda-nn). Without waiting for Cuda in the Render Pass that utilizes Cuda, the GPU time spent on Cuda spills to subsequent Render Passes, messing up the Profiling.

Or is there another possibility to profile CudaInterop in Falcor?

[julcst]

https://github.com/shader-slang/slang-rhi has support for external Semaphores on Vulkan it seems, so would I just need to update the slang version?