I am currently trying to compute vertices with CUDA and visualize them using Vulkan, following simpleVulkanMMAP in CUDA samples (via VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_BIT). But when I export a CUDA buffer with size larger than half of the device memory budget, ERROR_OUT_OF_DEVICE_MEMORY is reported by vkAllocateMemory.
The memory usage reported by VK_EXT_memory_budget and cudaMemGetInfo increases by the same amount of the imported CUDA memory after calling vkAllocateMemory, but the memory usage showed in windows task manager does not.
So is this a bug, or does importing memory from CUDA to Vulkan consume additional memory? Does Importing memory form Vulkan to CUDA have the same issue?
By the way, I am using RTX 3060 laptop and RTX 3090TI desktop (driver version 546.33)