[Offload] Name-based kernel loading

[e-kayrakli]

In implementing Chapel's GPU support, we have attempted to use libomptarget (soon-to-be llvm/offload?) in the past.

The main stumbling block was the expectation that the binary will contain a table of generated kernels. Our compiler currently doesn't do that, and our runtime rely on name based look up through the driver APIs such as cuModuleGetFunction/hipModuleGetFunction.

It'd be great if llvm/offload provide an interface for those functions.

[llvmbot]

@llvm/issue-subscribers-openmp

Author: Engin Kayraklioglu (e-kayrakli)

In implementing Chapel's GPU support, we have attempted to use libomptarget ([soon-to-be llvm/offload?](https://github.com//pull/75125)) in the past.

The main stumbling block was the expectation that the binary will contain a table of generated kernels. Our compiler currently doesn't do that, and our runtime rely on name based look up through the driver APIs such as cuModuleGetFunction/hipModuleGetFunction.

It'd be great if llvm/offload provide an interface for those functions.

[jhuber6]

I'm currently planning on making a new external API from the offloading library. This will be something more similar to using the CUDA driver or HSA and should hopefully make this process easier. I ran into similar issues while writing utilities in the libc project for loading kernels. Can't give you a definite timeline (as I'm on vacation for the rest of the month and my work capacity limited) but it's an active area of effort.

[jdoerfert]

Here is a POC to show how it could look like:
17aa944
(based on ed93e4a).

In my little test (not included) I could do:

  rc = __tgt_get_kernel_handle(0, argv[1], &Ptr);
  printf("rc: %i, ptr %p\n", rc, Ptr);

  rc = __tgt_launch_kernel_via_handle(0, Ptr);
  printf("rc: %i, ptr %p\n", rc, Ptr);

And launch all kernels in the device image, regardless if they were registered via our current mechanism or not.