tree 9b7425c90593a7a71cc8a56ba667988d476dff2a
parent fc0dd6ba6899bad16f10ec1d1964592ff1695b0b
author MaheshRavishankar <1663364+MaheshRavishankar@users.noreply.github.com> 1687890693 -0700
committer GitHub <noreply@github.com> 1687890693 -0700
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJkmysFCRBK7hj4Ov3rIwAAtQ0IAHsRVGpB1sCkb53OMjMPPEtF
 BEfDVA5pOb9V5P8jgRZPvQRcrkVgMlSaLAvVrsfqST1n3EvNkclbRNFjHMhHjCJX
 3Ux2vEprvtclQKmKYolAMDVqPUI7/W0VKM3lBTABtQ39BRY35ajBZuToxnhLSluZ
 /ANG6CGGDew3X/0TPEcB+i1CjPrIN3ft7Lg+jUfuHRjNIxzO3rCvDh6PTJNkuU+6
 Zj4o4zf0xU7FvGEQaLDII05Nr46DmDt8y+wfE5fLwkmmBAFjOXFXcSwcTrfSSrxW
 n818LsQHQtRwFUTIuqQ/XvvKvUAJYjCr3yzq+CFzQnWFaDtaVWISetJIZ6ZjQws=
 =Wf86
 -----END PGP SIGNATURE-----
 

Add support for lowering ops that implement the `UKernelOpInterface` on the CUDA path. (#14101)

This change add support for lowering operations that implement the
`UKernelOpInterface` on the CUDA backend. This allows custom
dispatches to use ops that implement and be compiled down to CUDA.

Unlike the CPU side, the CUDA side needs more information than just
the number of workgroups to use. It also needs the workgroup size. To
allow specifying this, an #iree_codegen.export_config attribute is
added. This allows specifying the workgroup size to use on the custom
dispatch. This information is used to set the workgroup size to use
during `MaterializeInterfaces` in HAL.

An example of the bitcode usage is added with this PR. Since there are
no in-tree examples of CUDA micro-kernels, the bit code is linked
using the `iree-link-bitcode` (which is the
`iree-llvmcpu-link-bitcode` renamed).

Fixes #14028