re above, i also noticed that dlpack already has an enumeration for webgpu and for vulkan. anyway. -
i'm looking a little bit into interoperability between vulkan and other things. using small smidges of human memory, so engaging avenues that kinda work. i'm presently on wsl, and the various development teams have things set up such that mesa works on wsl to render vulkan or opengl graphics, impressively. i downloaded the mesa sourcecode. it has backends for common accelerators (intel, metal, many other subfolders) for both opengl and vulkan. so i can look into the source and see directly what it is doing with the hardware pointers. intel for example has a canonical address that is exposed for vulkan buffers as the device address. if a dlpack interface were made, there's the device address right there. then the device itself would need to be enumerated and exported as well. now say i wanted to interoperate with torch. i'd have to worry about torch's backend; it's usually used with cuda and has had [creepily] more support for cuda than other things. but it also works with ati etc. here, if i'm on intel, i don't think that torch would accelerate on intel out of the box unless i downloaded some intel interface for it, which would kind of be a gamble since i'm in wsl, unlikely to work. so torch is going in software for me. but wsl is a pretty rare setup for accelerated computation on linux, i'm sure many people want it to work but it's kind of the thing that would be implemented last. ... but yeah dlpack could work in software by simply using the cpu device and the address. regarding nvidia, ati, this stuff would work too. regarding intel, it's not that common to use intel with ... but if i wanted accelerated computation, it might make sense to not use torch! i'm a little confused!
