intel
[SYCL][Graph][CUDA] Skip unsupported Windows E2E tests
On a CUDA & Windows setup when shared USM is used there is an issue with using the allocations concurrently in both device commands and host-tasks. This is based on an underlying CUDA restriction: https://forums.developer.nvidia.com/t/cudamallocmanaged-clarification-needed/67611
Applications running on Windows (whether in TCC or WDDM mode) or macOS will use the basic Unified Memory model as on pre-6.x architectures even when they are running on hardware with compute capability 6.x or higher.”
“Simultaneous access to managed memory on devices of compute capability lower than 6.x is not possible,”
Therefore, simultaneous access to managed memory on Windows is not possible.
This appears in SYCL-Graph tests where the graph has multiple roots, allowing host-tasks branching from one root to run concurrently with device commands from the other root. With the issue manifesting as a page fault in the host-task when trying to access a USM allocation. I've created a more minimal test test-e2e/USM/host-task.cpp which exhibits the same issues.
