[TensorIR] Pragma "import_llvm"

#1

Hi,

is there away to use the import_llvm pragma in TensorIR similarly how it can used in TE?

s[result].pragma(axis[0], "import_llvm", clang.create_llvm(cc_code, output=ll_path))

@junrushao @areusch @Hzfengsy

#2

I’m not sure but you can try to use

s.annotate(loop, "import_llvm", clang.create_llvm(cc_code, output=ll_path))

Genernally, it’s just an attr on the IR, so we should be able to do the same thing.

#3

if the goal is to inject some microkernels, tensorize is definitely the most canonical approach tho :slight_smile:

#4

I think this is s.annotate(loop, "pragma_import_llvm", clang.create_llvm(cc_code, output=ll_path))

There is a convention of prefix “pragma_” for te primitive.

1 Like
#5

Thanks. That’s correct. Just also figured it out :).

As @wrongtest, mentioned the codegen checks for "pragma_import_llvm"

Got it working like this

CC: @Hzfengsy @junrushao @SebastianBoblestETAS @UlrikHjort @aca88