Is it possible to use existing Vulkan context?

[trsh]

I know, vulkan is still in works, but would this be possible?

[Mike-Leo-Smith]

Yes. I can make the VulkanSwapchain interface class public so you can query the underlying VkInstance/VkPhysicalDevice/etc.

[Mike-Leo-Smith]

Ah, that would need some tricks to make the LC backends aware of the existing Vulkan objects. Maybe I can provide some options in DeviceConfig so that you can pass the objects when creating a Luisa device?

[Mike-Leo-Smith]

BTW, do you mean initializing the future Vulkan backend, or the existing VulkanSwapchain things in the CUDA/CPU backends?

[trsh]

Anything that makes LuisaCompute work. I just dont want to create second Vulkan context, devices etc. My main APP is already powered by vulkan, and its API exposes existing phy_device, device, vkGetInstanceProcAddr, vkGetDeviceProcAddr and vk_instance. This is how it looks in another product I am thinkin on using:

ctx = new VKContext(); // helper
ctx ->create(vk_instance, (void*)vkGetInstanceProcAddr_, vk_physical_device, vk_device, vk_family, 0) // Init with existing context features
otherFrameworkDevice = Device(ctx ) // create other frame work device, and use it as always - this could be Luisa device in our case orsmth

[Mike-Leo-Smith]

Understood. We'll consider this when making the Vulkan backend.

[Mike-Leo-Smith]

Users on other backends might also have the same demands. We should design a unified way of importing existing devices.

[trsh]

Users on other backends might also have the same demands. We should design a unified way of importing existing devices.

Yeah. In my case, I want to do some additional computation on GPU, and I don't want to go trough the hell of setting up 500 lines of vulkan code to run a compute kernel. Neither start new Vulkan context.